Essays on Applied EconometricsEssays on Applied Econometrics By Ebrahim Azimi Graduate Diploma,...
Transcript of Essays on Applied EconometricsEssays on Applied Econometrics By Ebrahim Azimi Graduate Diploma,...
Essays on Applied Econometrics
By
Ebrahim Azimi
Graduate Diploma, Institute for Advanced Studies, 2008
M.Sc., Institute for Management and Planning Studies, 2001
B.Sc. Amirkabir University of Technology, 1999
Thesis Submitted in Partial Fulfillment of the
Requirements for the Degree of
Doctor of Philosophy
in the
Department of Economics
Faculty of Arts and Social Sciences
Ebrahim Azimi 2016
SIMON FRASER UNIVERSITY
Spring 2016
ii
Approval
Name: Ebrahim Azimi
Degree: Doctor of Philosophy (Economics)
Title: Essays on Applied Econometrics
Examining Committee: Chair: Bertille Antoine Associate Professor, Department of Economics
Simon Woodcock Senior Supervisor Associate Professor, Department of Economics
Krishna Pendakur Co-Supervisor Professor, Department of Economics
Jane Friesen Supervisor Associate Professor, Department of Economics
Brian Krauth Supervisor Associate Professor, Department of Economics
Alexander Karaivanov Internal Examiner Associate Professor, Department of Economics
Guy Lacroix External Examiner Professor, Department of Economics Laval University
Date Defended/Approved: March 7th 2016
iii
Abstract
This thesis is composed of three essays on economics of labour, family, and education.
In the first chapter, I estimate the effect of having children on labour force participation of
mothers in urban Iranian areas. I exploit sex composition of children as an exogenous
source of variation in family size to account for endogeneity of fertility. Using information
from the Iranian Household Income and Expenditure Survey (HIES) over three samples,
namely households with one and more, two and more, and three and more children, I
find no significant effect of fertility on female labour force participation in Iran.
In the second chapter, I estimate family member’s resource shares and investigate
gender bias in intra-household resource allocation. I follow Dunbar et al. (2013) in that I
estimate the household member’s resource shares by observing how budget shares on
private assignable goods vary with total expenditure and family size. I extend their
methodology to analyze how sex composition of children influences resource shares.
Using data from the 2005 Iranian Household Income and Expenditure Survey (HIES), I
find that in Iranian rural areas parents assign 1.6 to 1.9% more resources toward their
sons. Similarly, I find that mothers in all-boy families get 2.8 to 3.6% less resources than
in all-girl families. These effects are more pronounced among farmer families. In
contrast, I find no significant role of gender composition on intra-household resource
allocation in Iranian urban areas.
In the final chapter I, jointly with Dr. Friesen and Dr. Woodcock, investigate the question
of whether schools that charge private tuition deliver higher quality education compared
to their public counterparts has proven very challenging. This paper contributes new
evidence regarding the quality of private schools relative to public schools. We use a
longitudinal student-level data set from British Columbia, Canada that comprises the
entire population of students in fourth through seventh grade who enrolled in public or
private schools. We apply a procedure developed by Abowd et al. (2002), which allows
us to exploit mobility between schools to estimate a full set of both school and student
fixed effects.
iv
Keywords: Female labour force participation, children sex composition, son preference, intra-household resource allocation, school quality, private schools, British Columbia
v
Dedication
To the loves of:
my wife, Maryam
and
my mom, Jannat
vi
Acknowledgements
I am deeply indebted to Professor Simon Woodcock for his continuous support, advice,
and encouragements. His wisdom made my Ph.D. studies a valuable journey. I am
especially grateful to Professor Krishna Pendakur for his valuable advice and stimulating
suggestions. The second chapter of this paper would not have been possible without his
precious help. I am also pleased to thank Professor Jane Friesen and Professor Brian
Krauth for their valuable time and support on my research and wonderful advice on my
teaching.
I would like to thank all faculties and staffs in Economics department at Simon Fraser
University especially Chris Muris, Andrew McGee, Hitoshi Shigeoka, Fernando Aragon,
Ramazan Gencay, and Gregory K. Dow for all wonderful discussions and their
comments.
I thank Mehdi Majbouri, Hossein Abbasi, Nick Dadson, Ideen Riahi, Natt Hongdilokkul,
all the participants of the labour and family session at the Canadian Economic
Association (2010) conference at University of Ottawa, participants at the 2011 HAND
forum at Massachusetts Institute of Technology, all participants of “International
Evidence on Empirical Microeconomics” session in 2014 CEA conference for their
helpful comments and discussions. All remaining mistakes are mine.
The administrative data used in the third chapter was extracted from the British
Columbia Ministry of Education’s student records by Maria Trache at Edudata Canada.
Klaus Edenhoffer created the digital maps used to link student postal codes to school
catchment areas using information provided by school district personnel. The programs
used to create the school choice variables used in the analysis were adapted from codes
written by Mohsen Javdani and Benjamin Harris. I appreciate their contribution to this
project.
vii
I acknowledge the financial support provided by Simon Fraser University’s Community
Trust Endowment Fund and the Social Sciences and Humanities Research Council of
Canada along this project.
viii
Table of Contents
Approval .......................................................................................................................... ii Abstract .......................................................................................................................... iii Dedication ....................................................................................................................... v Acknowledgements ........................................................................................................ vi Table of Contents .......................................................................................................... viii List of Tables ................................................................................................................... x List of Figures................................................................................................................. xi List of Acronyms ............................................................................................................ xii
Chapter 1. The Effect of Children on Female Labour Force Participation in Urban Iran ................................................................................................ 1
1.1. Data and Descriptive Statistics ............................................................................... 3 1.2. Children’s Sex Composition and Fertility ................................................................ 5 1.3. Estimation Results ................................................................................................ 12 1.4. Conclusion ............................................................................................................ 16
Chapter 2. Intra-household Resource Allocation and Gender Bias in Iran .......... 18 2.1. Iranian Context and Related Literature ................................................................. 20 2.2. Econometric Methodology .................................................................................... 23 2.3. Data ...................................................................................................................... 26 2.4. Results ................................................................................................................. 29
2.4.1. Estimates of Resource Shares ................................................................ 31 2.4.2. Children’s Gender Composition ............................................................... 35 2.4.3. Effect of Demographics ........................................................................... 41
2.5. Conclusions .......................................................................................................... 44
Chapter 3. Private Schools and Student Achievement ........................................ 46 3.1. Related literature .................................................................................................. 50 3.2. Institutional Context .............................................................................................. 51
3.2.1. Public school choice and funding ............................................................. 51 3.2.2. Private school choice and funding ........................................................... 52 3.2.3. Testing and accountability ....................................................................... 53
3.3. Data ...................................................................................................................... 54 3.4. Methodology ......................................................................................................... 55
3.4.1. The value-added model ........................................................................... 55 3.4.2. The student fixed effects model ............................................................... 57 3.4.3. Bounding the effects of selection on unobservables ................................ 59
3.5. RESULTS ............................................................................................................. 61 3.5.1. Descriptive statistics ................................................................................ 61 3.5.2. Results from the private school indicator model ....................................... 66 3.5.3. Results from the individual school fixed effects model ............................. 68
ix
3.6. CONCLUSION...................................................................................................... 77
References 80 Appendix A. Further Statistics ............................................................................... 87 Appendix B. Merging Student information to Census characteristics .................... 88
x
List of Tables
Table 1.1: Summary statistics: women aged 21–35 years ............................................... 6
Table 1.2: Fraction of families who had a second child depending on the sex of the first child ............................................................................................. 7
Table 1.3: Fraction of families who had another child depending on the .......................... 9
Table 1.4: Fraction of families who had a fourth child depending on Parity .................... 10
Table 1.5: Sex ratio by parity and sex composition of previous children ........................ 11
Table 1.6: Difference in mean for demographics by sex composition of children ........... 12
Table 1.7: OLS and 2SLS estimates of the effect of fertility on FLFP in urban Iran ........ 14
Table 2.1: Data Means .................................................................................................. 28
Table 2.2: Estimates of Parameters .............................................................................. 30
Table 2.3: Estimates of Resource Shares by Household Size ....................................... 32
Table 2.4: Estimates of Resource Shares by Household Size and Farmer Status ......... 34
Table 2.5: The Effect of Female Children on Family Members’ Resource Shares ......... 36
Table 2.6: The Effect of Female Children on Family Members’ Resource Shares ......... 38
Table 2.7: The Effect of Female Children on Family Members’ Resource Shares ......... 41
Table 2.8: The Effect of Demographics on Intra-household Resource Allocation ........... 42
Table 3.1: Selected school characteristics, by school type ............................................ 62
Table 3.2: Selected student characteristics, grade 7 students with non-missing test scores, by school type ..................................................................... 63
Table 3.3: Grade 7 student characteristics, by school type and mover status ................ 64
Table 3.4: Estimates of private school effect on reading and numeracy scores, value-added and student fixed effects models ........................................ 67
Table 3.5: Student-weighted means of estimated private school fixed effects on reading and numeracy scores, by type of private school, value-added and student fixed effects models ................................................. 69
Table 3.6: School-weighted means and standard deviations of estimated private school and student fixed effects on reading and numeracy scores, by type of school, student fixed effects model ........................................ 73
xi
List of Figures
Figure 1.1: Trend in Number of Children by Age Group and Family Size: 1990-2004 ......................................................................................................... 4
Figure 1.2: Trend in Female Labour Force Participation by Age Group and Family Size: 1990-2004 ....................................................................................... 5
Figure 3.1: Kernel density estimates of the distribution of estimated school effects, student FE model, full sample, by school sector ........................ 74
Figure 3.2: Kernel density estimates of the student-weighted distribution of estimated school effects, student FE model, full sample, by school sector ..................................................................................................... 75
Figure 3.3: Kernel density estimates of the distribution of estimated student effects, student FE model, full sample, by school sector ........................ 76
xii
List of Acronyms
FLFP
IV
HIES
Female Labour Force Participation
Instrumental Variable
Household Income and Expenditure Survey
1
Chapter 1. The Effect of Children on Female Labour Force Participation in Urban Iran
In economic literature, children are often considered a barrier to female labour
force participation (FLFP). In the last three decades, fertility in Iran has dropped sharply
from an average of seven births per woman in 1984 to less than two births in 2005.
Although fertility in Iran has experienced one of the fastest declines in modern human
history, no considerable rise in FLFP in Iran is documented (Aghajanian 1995; Abbasi-
Shavazi et al. 2009; Majbouri 2010). Using household-level information, this paper
investigates the effect of children on FLFP of mothers in urban Iran.
The association between fertility and FLFP is extensively documented in
theoretical models of work and family. While it is difficult to empirically estimate the
endogenous effect of fertility on FLFP (Schultz 1981; Goldine 1995), several studies
estimate its causal effect by exploiting an exogenous variation in family size. For
example, Rozenzweig and Wolpin (1980) and Bronars and Grogger (1994) use twinning
at the first birth. Angrist and Evans (1998) use an instrumental variables (IV) strategy
based on sibling sex composition in families with two or more children. Agüero and
Marks (2008) exploit random assignment of infertility as an exogenous variation in family
size. Using the Iranian Household Income and Expenditure Survey (HIES), this paper
contributes new evidence on the effect of fertility on FLFP by using an IV strategy.
I follow Angrist and Evans (1998) strategy to construct IV estimates of the effect
of fertility on FLFP based on sex composition of children. While in the US, parents are
2
more likely to have a third child if their first two children are of the same sex, in Iran, as
parents prefer sons to daughters, the presence of daughters in their previous children
acts as a positive shock to fertility. To show this relationship and investigate the effect of
children on FLFP, I construct three samples: one with families with one and more
children (1), another with two and more children ( 2
), and the third with three and more
children ( 3). In all these samples, families with more daughters than sons are more
likely to have another child. In other words, presence of more girls relative to boys
results in an increased likelihood of having another child. Considering this relationship, I
construct IV based on the sex composition of previous children to investigate the effect
of fertility on FLFP by using dummies for the gender of the first child, the first two
children, and the first three children in the samples of 1, 2
, and 3, respectively. As
sex-selective abortion and infanticide are rare in Iran, I consider sex composition among
children as essentially random. To support this claim, I follow Almond and Edlund (2008)
by observing that sex ratio does not vary significantly with birth order parity and sex
composition of the previous children.
To the best of my knowledge, this is the first estimation of the effect of fertility on
FLFP in Iran. While most empirical estimations of the effect of fertility on FLFP find a
negative impact, which in most cases is less negative than its ordinary least squares
(OLS) counterparts, I find no significant effect of children on the labour force participation
of Iranian mothers in urban areas. This result is similar to Agüero and Marks (2008), who
find an insignificant effect of fertility on FLFP in six Latin American countries.
The rest of this chapter is organized as follows. In the next section, I present the
data. In section 3, I describe the methodology and explain how fertility in Iran is
influenced by the sex composition of previous children. Section 4 presents the results,
and section 5 concludes.
3
1.1. Data and Descriptive Statistics
I use data from the HIES (1994–2003). This survey is conducted annually by the
Statistical Center of Iran. For each member of a family, this survey contains information
on demographic characteristics such as geographic location, age, gender, education,
relationship with the householder, marriage status, employment status, occupation, and
income. It also contains information on family expenditures, housing characteristics, and
ownership of assets and amenities.
The number of households in the HIES ranged from 17,500 in 1998 to 36,500 in
1995. To ensure that the insignificant effect of fertility on FLFP is not a result of
insufficient data, I use 10 rounds of the HIES data (1994–2003), all of which follow the
same definition for labour force participation.
The following restrictions are applied to the sample. 1) Polygamous families are
excluded from the sample; otherwise, it would have been impossible to match the
children to the women. 2) I exclude all cases wherein two or more families share a
common residence. Given that household identification, which is based on residential
address, is the only way to distinguish families, I am unable to distinguish children of
families that share the same residence. 3) Similar to most household surveys, the HIES
does not track children according to their households. To match the children with their
respective mothers, I restrict the sample to women aged between 20–35 years whose
oldest child is younger than 18 years. Few women younger than 19 have two or more
children, and women older than 35 are likely to have children who have already migrated
from the family. By restricting the sample, I ensure that the family’s oldest child is still
living with the parents and has not migrated from the family on account of marriage,
higher education, or work.
Table A1 in the appendix compares the selected sample with the overall sample
of women for some measures of fertility and FLFP. The three samples of women
include: (1) women aged between 20–35 years; (2) women aged between 36–50 years;
and (3) women aged between 20–35 years with two or more children and whose oldest
4
child is younger than 18. I highlight three features of fertility and FLFP in Iran from 1990–
2004 in this table. First, we observe a declining trend in the number of children; second,
depending on the year of the interview, a low FLFP rate of 10–14 percent in urban
areas, for the selected sample, is observed. The rigidity of FLFP in this period is an
important feature that has been addressed in the literature (Majbouri 2010). Third,
fertility and FLFP are comparable for all three categories.
Trends in fertility and FLFP are depicted in Figures 1 and 2, respectively, for
three subsamples of women from the HIES data.
Figure 1.1: Trend in Number of Children by Age Group and Family Size: 1990-2004
While fertility shows a sharp decline during this period, based on economic
theory, we expect an increase in FLFP. However, we observe no such increase over the
period in figure 2. The objective of this paper to investigate whether and to what extent
fertility impacts FLFP in urban Iran. I continue this discussion in section 4, where I
present the results.
1.5
2
2.5
3
3.5
4
4.5
1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004
Nu
mb
er o
f C
hild
ren
Year
21–35 year old women
36–50 year old women
21–35 year old women with 2+ children & age (oldest child > 18)
5
Figure 1.2: Trend in Female Labour Force Participation by Age Group and Family Size: 1990-2004
Table 1.1 reports the summary statistics for each sample of the 20–35 year-old
mothers in the three samples of 1, 2
, and 3.
1.2. Children’sSexCompositionandFertility
Similar to Angrist and Evans (1998), I use the following two-stage least squares
(2SLS) regression model of FLFP:
0 1= . .i i i ix w z ( 1.1)
( 1.2)
Here, ix is a measure of fertility for woman i ;
iy is an indicator of FLFP; iw
includes socioeconomic characteristics of i ; and iz denotes the instrumental variable
0.08
0.09
0.1
0.11
0.12
0.13
0.14
0.15
0.16
0.17
0.18
1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004
Fe
ma
le L
ab
or
Fo
rce
Pa
rtic
ipa
tio
n
Year
21–35 year old women
36–50 year old women
21–35 year old women with 2+ children & age (oldest child > 18)
6
based on the children’s sex composition. The instrumental variable is an indicator for the
children’s sex composition. The theoretical framework underlying the effect of son
preference on fertility is captured by the quality–quantity model of fertility developed by
Becker and Lewis (1973). Based on this model, in the current study, parents derive utility
from quantity and quality of children. Children sex composition in this model is viewed as
a source of utility related to the quality of children. If parents are not satisfied with the
sex composition of children, they will be more likely to expand their family to draw utility
from quantity of children or from a more desired composition of children. Therefore,
having more children is a response to dissatisfaction from children’s sex composition.
Table 1.1: Summary statistics: women aged 21–35 years
1 sample 2
sample 3
sample
mean St. dev. mean St. dev. mean St. dev.
FLFP 0.129 (0.335) 0.116 (0.320) 0.098 (0.297)
number of children 2.627 (1.431) 3.110 (1.279) 3.917 (1.131)
three-and-more-children indicator 0.447 (0.497) 0.579 (0.494) 1 (0)
four-and-more-children indicator 0.235 (0.424) 0.305 (0.460) 0.526 (0.499)
first-born son indicator 0.514 (0.500) 0.509 (0.500) 0.497 (0.500)
second-born son indicator 0.510 (0.500) 0.510 (0.500) 0.502 (0.500)
two-son indicator 0.260 (0.439) 0.260 (0.439) 0.255 (0.436)
two-daughter indicator 0.241 (0.428) 0.241 (0.428) 0.257 (0.437)
age 28.88 (3.967) 29.69 (3.685) 30.53 (3.376)
age at first birth 19.98 (3.462) 19.25 (3.096) 18.49 (2.745)
years of schooling 6.715 (4.267) 6.023 (4.076) 4.810 (3.727)
presence of relatives in the family 0.089 (0.285) 0.090 (0.286) 0.096 (0.295)
incidence of zero in non labour income 0.812 (0.391) 0.809 (0.393) 0.807 (0.394)
logarithm of non labour income 2.581 (5.426) 2.616 (5.449) 2.622 (5.427)
Observations 56845 43868 25399
Note: I report the mean of each variable with the standard deviation in parentheses. The variable “age at first birth” is measured by assuming that matching of the woman with their respected children is perfect. Therefore, this variable is measured by error.
7
Based on the above framework and similar to Ben-Porath and Welch (1976), I
analyze the effect of child sex composition on fertility among Iranian families. Table 1.2
reports how the firstborn’s sex influences the number of children. In the 1 sample, the
difference by the first child’s sex suggests that families with a firstborn daughter are one
percentage point more likely to have a second child. This is consistent with the
preference for a son among Iranian parents. Similarly, in the 2 and 3
samples, the
presence of a firstborn son reduces the likelihood of a third and a fourth child. Thus, the
firstborn’s sex is a plausible instrumental variable for number of children.
Similarly, among families with two or more children, the difference by the
firstborn’s sex suggests that those with a firstborn daughter are 2.4 percentage points
more likely to have a third child (table 1.3). This finding is also consistent with the fact
that Iranian parents have a marked son preference, especially for the first birth.
Table 1.2: Fraction of families who had a second child depending on the sex of the first child
1 sample 2
sample 3 sample
fraction of sample
fraction that had a 2nd child
fraction of sample
fraction that had a third child
fraction of sample
fraction that had a 4th child
(1) first-born son 51.3% 0.790 51.0% 0.603 50.6% 0.559
(0.002) (0.003) (0.004)
(2)first-born daughter 48.7% 0.799 49.0% 0.627 49.4% 0.583
(0.002) (0.003) (0.004)
Difference (1)-(2) -0.009*** -0.024*** -0.024***
(0.003) (0.004) (0.005)
Note: Standard Errors are reported in parenthesis. ***: significant at 1% level.
8
Similar to the case of the firstborn daughter, the second daughter also increases
the likelihood of having a third child. The effect, however, is smaller than that of the
firstborn girl. Parents of a second daughter are 1.7 percentage points more likely to have
a third child. Further, while parents with two sons are 1.2 percentage points less likely to
have a third child, parents with two girls are 4.5 percentage points more likely to have a
third child. This also shows that parents of mixed sex children are 2.4 percentage points
less likely to have a third child relative to parents of same-sex children.
There are three points worth mentioning about table 1.3. First, all the evidence
demonstrates a marked preference for sons among Iranian families; while sons reduce
the likelihood of a third child, daughters increase it. Second, the effect of two daughters
is larger compared to other combinations. Third, this table shows that child sex
composition has a strong explanatory power on fertility decisions in Iranian families.
Based on these results, I use an indicator of two daughters as an instrumental variable
to estimate the effect of a third child on FLFP in the 2 sample.
Similarly, table 1.4 reports the relation between sex composition of children and
probability of having a fourth child in families with at least three children. Based on this
table, I use the indicator of having at least two sons as an instrument to investigate the
effect of a fourth child on FLFP in the 3 sample.
Random assignment of sex composition makes it very likely that these IV
estimates of the effect of fertility on FLFP have a causal interpretation. Selective abortion
and infanticide are quite rare in Iran because both have been illegal since the 1979
Islamic Revolution, except in very specific cases where the mother is in serious danger
or the baby is expected to born with a severe disease (Hoodfar 1996; Mehryar et al.
2007). Thus, we treat the gender composition of children as essentially random.
9
Table 1.3: Fraction of families who had another child depending on the
sex composition of previous children: 2 sample
fraction of sample
fraction that had a third child
(1) first-born son 51.0%
0.603
(0.003)
(2) first-born daughter 49.0%
0.627
(0.003)
Difference: (1)-(2)
-0.024***
(0.004)
(1) second-born son 51.1%
0.607
(0.003)
(2) second-born daughter 48.9%
0.624
(0.003)
Difference: (1)-(2)
-0.017***
(0.004)
(1): two sons 26.2%
0.607
(0.004)
(2):not(two sons) 73.8%
0.618
(0.002)
Difference: (1)-(2)
-0.012***
(0.005)
(1): two daughters 24.0%
0.649
(0.004)
(2): not (two daughters) 76.0%
0.604
(0.002)
Difference: (1)-(2)
0.045***
(0.005)
(1): mixed sex 50.2%
0.627
(0.003)
(2): same sex 49.8%
0.603
(0.003)
Difference: (1)-(2)
0.024***
(0.004)
Note: Standard Errors are reported in parenthesis.
***: significant at 1% level.
10
Table 1.4: Fraction of families who had a fourth child depending on Parity
and sex composition of previous children
fraction of sample
fraction that had a 4th child
(1): 3 sons 13.6% 0.560
(0.007)
(2): other mixes 86.4% 0.573
(0.003)
Difference: (1)-(2) -0.013**
(0.007)
(1): 3 daughters 12.7% 0.557
(0.007)
(2): other mixes 87.3% 0.573
(0.003)
Difference: (1)-(2) -0.015**
(0.008)
(1): same sex 25.8% 0.594
(0.005)
(2): mixed sex 74.2% 0.563
(0.003)
Difference: (1)-(2) 0.031***
(0.006)
(1): 2 sons, 1 daughter 38.0% 0.548
(0.004)
(2): other children compositions 62.0% 0.585
(0.003)
Difference: (1)-(2) -0.036***
(0.005)
(1): 1 son, 2 daughters 36.2% 0.578
(0.004)
(2): other children compositions 63.8% 0.567
(0.003)
Difference: (1)-(2) 0.011**
(0.005)
Note: Standard Errors are reported in parenthesis. ***: significant at 1% level, **: significant at 5% level.
11
One empirical test for random assignment of child sex composition is to compare
the sex ratio by birth order and sex of previous children (Almond and Edlund 2008). For
the general human population, sex ratio, defined as the proportion of males to females,
is about 1.05 with the exception of during and after war times, where it is documented to
be slightly higher. Otherwise, a higher-than-normal sex ratio is a sign of sex-selective
abortion. In such cases, the sex ratio will depend on previous children’s sex composition.
For example, due to the availability of prenatal sex determination, a two-child family that
already has a daughter and prefers sons is more likely to abort a girl fetus relative to a
boy fetus. Table 1.5 reports sex ratio by birth order and previous children’s sex
composition. No significant inflated sex ratio is evident from this table. Although the sex
ratio is slightly larger than 1.05 for the third child, there is no significant difference
between the sex ratios of families with two sons and families with two daughters. Thus,
child sex composition can be treated as random. The results of the 1996 and 2006
Iranian Censuses in table A2 in the appendix support the randomness of sex
composition in Iran.
Table 1.5: Sex ratio by parity and sex composition of previous children
birth previous
observations
sex 95% confidence interval
order children ratio lower bound upper bound
first 132,599 1.055 1.044 1.066
second girl 51,689 1.032 1.015 1.050
boy 53,920 1.049 1.031 1.066
third girl, girl 17,356 1.082 1.051 1.115
girl, boy 33,703 1.067 1.044 1.090
boy, boy 17,685 1.080 1.048 1.112
An alternative test for random assignment of child sex composition is to compare
demographic characteristics of families by the sex composition of children (Angrist et al.
1998). If child sex composition is random, there is no significant difference between
12
demographic characteristics of families by this composition, as the insignificant
differences in table 1.6 suggest.
Table 1.6: Difference in mean for demographics by sex composition of children
a first-born son two daughters two or more sons
age -0.012 0.030 -0.020
(0.026) (0.034) (0.036)
literacy -0.002 -0.004 0.006
(0.003) (0.005) (0.006)
education -0.013 -0.050 0.051
(0.035) (0.044) (0.044)
husband’s years of education -0.004 -0.042 0.081
(0.038) (0.049) (0.053)
Sample 1 sample 2
sample 3 sample
Note: Standard errors are reported in parentheses.
The random assignment of child sex composition makes it very likely that the
reduced form regressions of fertility and FLFP have a causal interpretation. Section 4
reports this estimation, and the results confirm that the three dummies of firstborn son,
two daughters, and two sons and more are plausible instrumental variables for
investigating the effect of fertility on FLFP in the 1
, 2, and 3
samples, respectively. I
use number of children and indicators of having more than two and three children as the
measures of fertility for the 1
, 2, and 3
samples, respectively.
1.3. Estimation Results
In this section, I estimate the effect of having more children on FLFP, using the
2SLS regression model of FLFP shown in equations (1.1) and (1.2). As explained
earlier, I estimate the model for three subsamples of women. I use number of children,
13
an indicator denoting that a woman has more than 2 children, and indicator denoting that
a woman has more than three children as measures of fertility for the samples 1, 2
,
and 3
, respectively. The respective instrumental variables are indicators of a firstborn
son, two daughters, and two or more sons.
All the specifications include indicators of age, namely, I(25 ≤ age ≤ 29), I(30 ≤
age ≤ 35); indicators of schooling, namely, I(1 ≤ schooling ≤ 5), I(6 ≤ schooling ≤ 8), I(9 ≤
schooling ≤ 12), I(13 ≤ schooling); age at first birth, log(nonlabour income), and year
effects.
Table 1.7 reports the results of the OLS and IV estimates of the effect of children
on FLFP for all three samples as well as the results of the first-stage estimates.
All three samples confirm the strong association between fertility and child sex
composition. The F-statistics of a test for weak instrument hypothesis based on Stock
and Yogo (2005) strongly rejects the null hypothesis of a weak instrument. The IV
estimates suggest no significant effect of fertility on FLFP, while the OLS estimates
report negative effects. Although the OLS estimates are small, they are significant. Most
studies of the causal effect of fertility on FLFP find a negative effect, which is usually
smaller than the OLS estimates but is still significant. For example, using the firstborn’s
sex as an instrumental variable; Chun et al. (2002) find that an additional child reduces
the labour force participation of Korean mothers by 27%. Angrist et al. (1988) estimate
the effect of a third child on women’s income to be -0.12. On the other hand, Agüero and
Marks (2008) find no evidence that fertility has a causal effect on FLFP.
Although the insignificant effect of fertility on FLFP is consistent with the
aggregate trend of FLFP for urban Iranian families (see Figures 1 and 2), the result is
nevertheless surprising.
As explained in section 2, I restrict the sample according to the ages of the
women and their oldest child, to match the women with their respective children. If the
match is not close to perfect, the estimate of the variable “age at first birth” will be
14
erroneous resulting in bias in estimates. In my result, however, excluding this variable
does not change the main finding of the paper; that is, fertility continues to have an
insignificant effect on FLFP.
Table 1.7: OLS and 2SLS estimates of the effect of fertility on FLFP in urban Iran
1 sample 2
sample 3 sample
OLS 2SLS OLS 2SLS OLS 2SLS
First-stage results: fertility equation
a first-born son -0.106***
(0.009)
two daughters 0.042***
(0.004)
two or more sons -0.044***
(0.005)
Estimation Results: FLFP equation
number of children -0.008*** 0.005
(0.001) (0.026)
more than 2 children -0.011*** -0.033
(0.004) (0.060)
more than 3 children -0.009** -0.019
(0.004) (0.068)
Observations 56,845 56,845 43,868 43,868 25,399 25,399
1st stage R-squared 0.562 0.344 0.226
2nd stage R-squared 0.183 0.183 0.152 0.151 0.084 0.083
IV F-statistics 149.253 149.430 89.917
Note: All the specifications include indicators of age: I(25≤age≤29), I(30≤age≤35); indicators of schooling: I(1≤ schooling≤ 5),I(6≤ schooling ≤8),I(9≤ schooling≤ 12), I( schooling≤ 13); age at first birth, log(non-labour income), incidence of zero in non-labour income and year effects. Standard errors are reported in parentheses. ***: significant at 1% level, **: significant at 5% level.
One potential concern with the estimation of FLFP equation is endogeneity of
nonlabour income as it highly depends on previous labour income (Lise and Seitz,
2011). The endogeneity of income can potentially create a bias to the estimated effect of
children on FLFP. To control for this endogeneity problem, one potential instrument is
changes in liquidity such as changes in local housing market prices (Hurst and Lusardi,
2004). In this study, however, using similar instrument is not feasible with data
15
restriction. Therefore, as a robustness check, I estimate the FLFP equation by excluding
the non-labour income. No effect of children on FLFP is robust to the omission of
nonlabour income which reduces concerns about the endogeneity of nonlabour income
in this study.
My finding suggests that children at the extensive margin are not a barrier for the
female labour supply. However, one possibility is that fertility influence labour supply of
women at the intensive margin. That is, women may change their work hours in
response to having a larger family. Unfortunately, information on hours of work are not
available in the HIES. Therefore, it is not feasible to estimate the effect of children on
female labour supply at the intensive margin. Even if this is a plausible explanation, the
rigidity of FLFP at low levels is still surprising.
The low rate of FLFP in Iran and its rigidity over the last three decades have
been referred to as a puzzle in development literature (Majbouri 2010). It is more
surprising to know that this rigidity was concurrent with a sharp decline in fertility, as
mentioned earlier, and a considerable increase in the education of women. For example,
according to the Statistical Center of Iran, the female-to-male student ratio in Iranian
public and private colleges has increased from less than 0.4 in 1990 to about 1.2 in
2005.
With the sharp decline in fertility and the considerable increase in women’s
education in Iran, we might expect an increase in FLFP. However, FLFP has remained
low at 10–14 percent. While investigating the reasons behind the low rate and rigidity of
FLFP is outside the scope of this paper, I state a few potential reasons. One possible
explanation is the hidden employment resulting from the definition of employment in
governmental surveys that measure FLFP including HIES. These surveys consider
someone employed only if they work at least two days per week. This may preclude
people employed on a part-time basis. Another possible explanation is that traditional
norms stemming from religion have a strong role in impeding women from seeking
employment. In addition, employment of women can increase their autonomy within the
family and it reduces the autonomy of their husband. The fact that women employment
16
influences the intra-household bargaining power provides further explanation on barriers
to FLFP in Iran.
The recent development literature provides some additional possible
explanations for low FLFP in the region and its rigidity in the last few decades. Many
observers believe that religion is the main reason for the low representation of women in
the labour market (Sharabi 1988). They emphasize the common problem of low FLFP
throughout the Middle East and attribute it to the influence of traditional and religious
norms. The suggested mechanism is that traditional norms result in discrimination
against women in the labour market, by restricting both labour supply and labour
demand for women. This explanation, however, does not seem plausible as countries
like Bangladesh and Indonesia are predominantly Muslim but have high rates of FLFP.
Using cross-country data, Ross (2008) shows that the effect of Islam disappears
as he controls for oil and gas income, and he concludes that oil, not Islam, is responsible
for low rates of FLFP in Iran and other Middle Eastern countries. Majbouri (2015)
challenges this argument by proposing a mechanism through which oil and gas income
along with traditional institutions account for the rigidity of FLFP. He explains that oil and
gas income acts as rent and strengthens traditional norms and the religion’s influence
among Muslim countries with access to oil income.
1.4. Conclusion
Rigidity of FLFP in urban Iran in the last three decades has been a consensus in
development literature. It is of more surprise to know that it has been simultaneous with
the period in which fertility has sharply declined and women’s education has
considerably increased (Majbouri 2010). I investigate the causal effect of fertility on
FLFP to shed light on a part of this puzzle.
Following Angrist et al. (1988), I exploit the random assignment of child sex
composition as an instrument to investigate the causal effect of fertility on FLFP among
Iranian families in urban areas. As Iranian parents prefer to have sons relative to
17
daughters, I show that the presence of sons reduces the likelihood of having more
children in Iranian families. Based on this information, I exploit children’s sex
composition to investigate the effect of fertility on FLFP.
While most estimates of the causal effect of fertility on FLFP report negative
effects, I find no evidence that the presence of more children is a barrier for mothers to
work. This finding is similar to that of Agüero and Marks (2008) and consistent with the
aggregate trend in FLFP in urban Iran. Oil and gas income along with traditional
institutions in Iran are considered to account for the rigidity of FLFP (Majbouri, 2015).
18
Chapter 2. Intra-household Resource Allocation and Gender Bias in Iran
Intra-household resource allocation has been widely studied in analyzes of
gender bias, poverty, and standards of living. The gender gap observed in adult
outcomes is generally thought to be a consequence of gender bias in intra-household
resource allocation (Deaton, 1997). While preference for sons as well as the gender gap
in a number of socioeconomic has been reported in Iran, intra-household resource
allocation and its gender bias consequences remain open questions. Given the
foregoing, I estimate the resource share allocated to men, women, and children among
Iranian families in rural and urban areas.
In particular, I investigate two potential consequences of son preference1 on
intra-household resource allocation in Iran. First, I examine whether and to what extent
son preference among Iranian parents results in the allocation of family resources in
favor of boys relative to girls. Second, I investigate the effect of family gender
composition on the resource share of parents. If parents prefer boys relative to girls, the
trade-off between their own consumption and their children’s consumption is expected to
differ for boys and girls; specifically, there would be relatively less consumption for
parents when they have sons. Furthermore, I expect the extent of this trade-off to differ
between parents. As Iranian culture is strongly patriarchal ( Moghadam, 1992), the effect
of more sons in a family on the mother’s resource share is of great scholarly importance.
__________________________________ 1 Son preference refers to the attitude based on which a female child is less valued than a male child by her parents and the society.
19
The effect of patriarchy on women’s resource share is ambiguous because at
least two opposing forces exist. First, since giving birth to a son is a source of pride and
privilege for a woman, the presence of more boys in the family leads to more autonomy
for the mother, resulting in her higher resource share. In this case, the father makes the
largest sacrifice by allocating more of his own resources to his son. Second, the
presence of boys in the family may create competition between the mother and her
sons, resulting in a relatively low resource share for the mother. While both views are
similarly plausible, I investigate which one is consistent with household behavior in Iran.
This study provides the first estimates of intra-household resource allocation in
Iran, and quantifies the effect of gender composition on resources allocated to each
family member among Iranian families. This has two principle implications in welfare
economics. First, it sheds light on a major source of economic inefficiency, especially in
developing countries. Gender bias in intra-household resource allocation is a source of
discrimination that could lead to inequality in adulthood by negatively influencing the
well-being of female offspring and promoting a gap in the educational and labour market
outcomes of girls compared with boys. Second, the estimation of the resource shares of
household members helps measure well-being at an individual level. If households
unevenly allocate resources among family members, then a household-level analysis of
poverty, for example, might mask the effect of adverse economic conditions on individual
family members.
Unitary models of households are unable to examine the within-household
distribution of consumption (i.e., the resource shares of household members). By
contrast, collective household models (e.g., Chiappori, 1988, 1992), wherein the
household is modeled as a collection of individual people with distinctly defined utility
functions, are more suitable for assessing distribution among household members. The
main contribution of this study is thus using a collective household model to investigate
empirically gender bias in intra-household resource allocation.
In general, the effect of gender composition on the resource allocation of Iranian
families is difficult to study because household surveys usually provide expenditure
information at the household level and not at the individual level. This study uses the
collective household model of Dunbar et al. (2013) as well as information on assignable
goods in household-level expenditure data plus a structural model of demand to
estimate the share of household resources allocated to each family member. The
approach of Dunbar et al. (2013) identifies resource shares by observing how family
20
expenditure on private assignable goods such as clothing varies by household type and
total expenditure. I use the Iranian Household Income and Expenditure Survey (HIES)
2005, and, like Dunbar et al., use clothing as the private assignable good. But, in
contrast to Dunbar et al., I focus on how resource shares depend on gender composition
of the family.
The estimated results, the first on intra-household resource allocation in Iran to
my best knowledge, suggest that Iranian families in rural areas allocate more resources
toward their sons. Depending on family size, Iranian parents in rural areas devote 15–
17% of their resources toward boys. The share of girls, however, is on average 1.6–1.9
percentage points less than that of boys. In addition, the presence of boys in the family
diverts resources from their mothers. In all-boy families, mothers obtain fewer resources
than those in all-girl families, by 2.8–3.6 percentage points. By contrast, the results from
families in urban areas suggest no strong effect of gender composition on household
members’ resource shares.
The remainder of the paper proceeds as follows. Section 2 reviews the related
literature. Section 3 summarizes the econometric methodology. Section 4 describes the
data set. Section 5 presents and discusses the empirical results. Section 6 concludes.
2.1. Iranian Context and Related Literature
Studies of economic well-being are of particular interest for Iranian households
because the country has experienced three decades of adverse economic conditions.
The revolution of 1979, the 1980–1988 war with Iraq, and international economic
sanctions have all negatively influenced Iranian families’ standards of living. In addition,
considering the patriarchal culture among Iranian families, it is important to account for
within-family inequality when measuring individual well-being, especially in developing
countries such as Iran where family decisions play a dominant role in children’s future
options. In Iran, for example, families take on the roles assigned to other social
institutions in more developed countries. In this regard, Iranian households decide
whether and to what extent a child should be educated. By contrast, because the law in
developed countries enforces minimum education levels, there is less potential for
discrimination against girls.
21
If resource allocation in Iranian families is discriminatory against female offspring,
then it may have wide-ranging influences on many dimensions of their lives. Indeed,
gender bias in intra-household resource allocation could explain the differences in many
socioeconomic outcomes, such as the higher mortality rate of girls relative to boys
(UNFPA, 2010), their lower secondary school enrollment rate (UNICEF, 2007), and poor
female labour market outcomes (Abbasi and Farjadi, 1999). Indeed, Azimi (2015) finds
that the gender composition of children even influences Iranian parents’ decisions
regarding fertility.
Detecting gender bias in intra-household resource allocation is a challenging
task. If consumption were observed at the individual level, one could directly observe
whether some members received a greater share of household resources than others.
Unfortunately, household consumption surveys usually provide information on various
types of commodity expenditures at the household level not at the individual level.
Therefore, drawing conclusions from these data about gender bias at the household
member level is not straightforward.
Some of the stream of research on intra-household resource allocation uses
information on aggregate consumption (e.g., using expenditure on food, education, or
healthcare as a proxy) and family composition to investigate how household expenditure
needs depend on the age and gender composition of families. For example, Deaton
(1997) uses a regression of food budget share on the natural logarithm of family size,
the natural logarithm of per capita total expenditure, the proportion of family members in
each age–sex group, and household demographics. The share of food in family total
expenditure is considered as an inverse measure of household welfare. The coefficient
of each age–sex group identifies the increase in expenditure share of the analyzed good
as a result of adding one more member in that group. Although researchers have used
this method to analyze a number of countries including India, Bangladesh, Pakistan, and
China (Subramanian and Deaton, 1991; Bhalotra and Attfield, 1998; Gong et al., 2000),
establishing that household consumption needs are related to the gender composition of
families; this finding does not imply that consumption is shared unequally within the
household.
Although the analysis of household-level food, health, or educational
consumption is not in general informative about discrimination in the intra-household
resource allocation, the household-level consumption of exclusive goods is informative
in this regard. An exclusive good is one consumed by a single identifiable household
22
member. Such goods are thus useful because the household-level consumption of the
good equals the individual-level consumption. The analysis of exclusive goods is a
common approach for investigating gender bias in intra-household resource allocation.
In this context, however, we need goods exclusively consumed by different genders (and
potentially ages).
Rothbarth (1943) first applied this concept to observe how the slope of the Engel
curve for an adult good such as alcohol, tobacco, or adult clothing varies with the gender
of children. In this case, the consumption of adult goods is a proxy for parents’ welfare.
The slope of the Engel curve shows the extent to which adults are willing to trade off
their own welfare as well as that of their children. If the extent of this trade-off is different
for boys and girls, it is a sign of gender bias. While this approach originally estimated the
cost of children, researchers have since used it extensively to estimate intra-household
resource allocation. Gronau (1988) shows that the theoretical basis for the Rothbarth
methodology is the assumption of separability and constant preferences. Alcohol,
tobacco, clothing, and footwear are among the exclusive goods used in the literature.
This method and its extensions have been widely used for intra-household resource
allocation studies in India and China (Lancaster et al. 2003; Gong et al. 2000; Kingdon
2005; Zimmerman 2012).
Two main criticisms arise from the assumptions of previous studies on the
household-level consumption of shared or exclusive goods. First, earlier research has
typically used a unitary model of households wherein the household optimizes a single
utility function rather than an amalgam of each member’s utility function. Investigating
intra-household resource allocation and gender bias based on a model that comprises
one utility function (i.e., only one person) seems to be unsuitable. Second, these
methods do not account for returns to scale in consumption. Adding the same resources
to families of different sizes can result in different welfare. Since large families usually
benefit from economies of scale in consumption, research on intra-household resource
allocation should overcome this discrepancy.
This study avoids these drawbacks by following Browning et al. (2013) and
Dunbar et al. (2013), who propose collective households (i.e., they model the household
as a collection of members with individual utility functions) that allow for scale economies
in consumption. In particular, I adapt those models to address the possibility that the
intra-household distribution of resources depends on the gender composition of children.
23
2.2. Econometric Methodology
The collective model of Dunbar et al. (2013), based on that of Browning et al.
(2013), proceeds as follows. Households consist of three types of individuals t={f,m,c},
denoting the father, mother, and children, respectively. The number of children
s={0,1,2,3} is used to index household type. Let y denote the household’s total
expenditure and p denote the market price vector.
Household consumption decisions take into account the utilities of each person;
hence, I assume that such decisions reach the Pareto frontier in the sense that trade
across household members does not lead to unexploited gains. This assumption allows
me to decentralize the household problem into a separate optimization problem for each
member. Separability assumption also allows for separability of consumption and
production.
The personal optimization problem aims to maximize utility against a budget
constraint characterized by a shadow price vector r (which is the same for all household
members) and a shadow budget ηts
y.
The shadow price vector r does not equal the market price vector p because
shareable goods have shadow prices below their market prices. These shadow prices
capture what Browning et al. (2013) call the “consumption technology,” while the
differences between the shadow and market prices account for scale economies in
household consumption. Importantly, the shadow price of an exclusive good is equal to
its market price, because exclusive goods are by definition not shareable.
The scalar ηts
is the resource share of person t in a household with s children.
The resource share is a scalar-valued measure of an individual budget constraint within
the household, and it measures the amount of consumption flowing to that person (i.e.,
the object of interest in this paper).
Define Wts
(y) as the share of a household’s total expenditure spent by member t
on his or her private assignable good in a household of type s. Let wt(y) be the
hypothetical share of y that t would spend on his or her private good when maximizing
his or her own utility function subject to shadow price r and budget y. In general, Dunbar
24
et al. (2013) show that if the data have no single-member households, further structure
is necessary to identify the resource shares and thus they propose the restriction that
resource shares are independent of y and a pair of preference restrictions, which allows
them to identify resource shares from the variation in the Engel curve.
These authors also specify the solution to this household optimization problem in
the form of an Engel curve relating budget shares Wts
(y) to budgets y, as follows:
(2.1)
In Eq. (2.1), Wts
(y) is observable from the information available on household
expenditure information. Here, I assume that ηts
does not depend on y. The goal is to
identify the resource share in Eq. (2.1). However, the challenge in identifying the
resource share in this equation is that for every observable Wts
(y) on the left-hand side,
two unknown functions of resource share exist on the right-hand side, namely ηts
and
wt(η
tsy). This is where the preference restrictions come in.
Dunbar et al. (2013) impose that the functions wt(η
tsy) have similar shapes
(essentially fixed curvatures) for either variation in household size s or variation in
person t. Under this structure, no further restriction on the shape of the preference
functions wt(η
tsy) is necessary to identify the resource shares.
For estimation simplicity, let preferences be price-independent generalized
logarithmic (PIGLOG; Muellbauer, 1976). In this case, demands are given by
(2.2)
25
These two restrictions are when preferences are similar across people (SAP) or
when preferences are similar across types (SAT). The SAP condition for a PIGLOG
indirect utility function is equivalent to βts
=βs for all individual types. Under SAP, the
linear regression of the private good’s household budget shares on ln(y) allow us to
identify the slopes of the abovementioned household Engel curves. Considering that
resource shares sum to one for each household type, we can identify three resource
shares and βs. Alternatively, the SAT condition for a PIGLOG indirect utility function is
equivalent to βts
=βt for all household and individual types. Considering four household
types, Eq. (2.3) provides 12 equations in addition to the three sets of resource shares
equal to one. This enables us to identify the three resource shares for each household
type plus the three preference parameters, βt. With more than four household types, the
model is overidentified.
We can use additional information to test the model or improve the precision of
the estimates. Dunbar et al. (2013) combine both restrictions by imposing βts
=β. In this
study, I use information on Iranian families with zero to three children. The presence of
childless families makes the SAT restriction a strong assumption. Therefore, I estimate
the parameters by assuming that a combination of SAT and SAP holds for families with
one to three children (β is allowed to differ for childless families). This assumption can be
stated as βts
=β1 for s=1,2,3 and β
ts=β
0 for s=0. The corresponding household Engel
curves for the private assignable goods under the assumption of a combination of SAP
and SAT are
(2.3)
where β=(β1*I
s>0+β
0*I
s=0).
26
I use the following parametric specification that allows resource shares to depend
on household size and the gender composition of children:
(2.4)
This specification consists of a quadratic function of family size multiplied by a
linear function of the proportion of girls in the family. α0t
,α1t
,α2t
are coefficients of the
quadratic term, while sg
is the number of girls and γt the coefficient of the proportion of
girls for each family member, t. This specification has a number of useful features. First,
the quadratic functional form allows returns to scale in consumption. Second, the
intercept in the female and male equations identifies the resource share in a childless
family. Third, the children equation does not include an intercept because children’s
resource share in childless families is zero. Finally, the extent to which each member is
influenced by the presence of female children in the family is proportional to γt.
To estimate this model, I add an error term to each equation and use nonlinear
seemingly unrelated regression techniques to estimate the resource shares on a sample
of Iranian households with zero to three children.
2.3. Data
The HIES is an annual survey conducted by the Statistical Center of Iran. For
each family member, the HIES reports demographic information such as geographical
location, age, gender, education, relationship with the family head, marital status,
employment status, occupation, income, and sources of income. The focus of this
survey is, however, on household expenditure, for which the HIES reports information on
about 600 items. Expenditure on clothing and footwear is reported separately for male
adults, female adults, and children in different age groups. I combine the children age
27
groups to construct clothing and footwear expenditure for children younger than 14
years. The recall period for consumption expenditure is one month. To capture seasonal
effects, households are interviewed in different seasons (the interview season is
reported). I use information on interview season to test the robustness of results when I
account for any price variation over a year.
The HIES included about 24,000 families in 2005. I restrict the sample to
childless couples and families composed of married couples with one to three children.
Furthermore, I exclude polygamous families, households with any member older than 65
years, two or more families sharing a common residence2, and families with children
older than 14 years. Finally, I limit the sample to families comprising a father, mother,
and zero to three children (i.e., I omit families that include other relatives such as
grandparents). The final sample includes 7,496 households. While it is technically
difficult to measure the effect of the above sample restrictions, it is unlikely that these
restrictions change the main results. In section 5, however, I briefly explain a potential
consequence of the sample restrictions.
I use clothing as the exclusive good for adults. The exclusion of families with
children 14 or over allows us to assign clothing expenditure to either adults or children,
and since the data separate clothing expenditure for men and women, to husbands and
wives3. For children, I create a children’s good that includes both clothing and other
children’s private goods such as toys, dolls, and children’s clothing. Table 2.1 presents
the means of the exclusive goods for different family members, total household
expenditure, and demographic characteristics for the sample by family size.
Further, a rich set of demographic variables from the HIES is used in this study:
an indicator of urban–rural residential area, age of father, age of mother, average age of
children, age of the youngest child, father’s education, mother’s education, indicator of
parents’ occupation in farms, and indicators of the interview season.
The inclusion of demographics, although not required to identify resource shares,
helps to identify the share more precisely. Hence, I estimate the resource shares for a
reference family (i.e., a family for which all the demographics have zero values). For all
__________________________________ 2 Given that household identification (based on residential address) is the only way in which to distinguish families, I am unable to decompose those families that share the same residence.
3 Because four-child families in the sample were rare, I restrict the sample to couples with zero to three children.
28
age and educational variables where zero values are rare or impossible, I include the
deviation from the modal values. Specifically, I measure the ages of the father and
mother as deviations from 35 and 30 years, respectively and the average age of children
and age of the youngest child as deviations from five. Similarly, I measure parents’
education levels as deviations from five. Five years of schooling is equivalent to
completing primary school, which is the most common level of education among parents
in the sample. Using deviation from the mode compared with deviation from the mean
ensures a sufficient proportion of reference families.
Table 2.1: Data Means
couples with
no child 1 child 2 children 3 children 0-3 children
men’s clothing and footwear 0.028 0.022 0.019 0.018 0.021
women’s clothing and footwear 0.023 0.018 0.014 0.012 0.016
children’s private good 0 0.035 0.042 0.047 0.033
number of children 0 1 2 3 1.45
number of daughters 0 0.48 0.98 1.46 0.71
log(total expenditure) 14.5 14.6 14.7 14.6 14.6
men’s age(demean) 1.58 -2.81 -0.45 0.73 -0.67
women’s age(demean) 1.42 -2.61 -0.24 1.05 -0.51
men’s schooling (demode) 2.98 3.44 2.61 1.19 2.75
women’s schooling (demode) 2.25 3.10 1.75 -0.38 1.98
average age of children -5 -0.47 1.21 2.04 -0.35
children age difference 0 0 3.98 6.90 2.38
first born girl indicator - 0.48 0.51 0.49 -
proportion of girls - 0.48 0.49 0.49 -
Observations 1377 2379 2718 1022 7496
All demographics, including age of father, age of mother, average age of
children, children’s age difference, father’s education, and mother’s education, enter the
model through linear shifters to the resource share functions ηts
and through linear
shifters of each member’s preferences through δfs
. I use all other information for
robustness checks.
29
2.4. Results
I first present the coefficient estimates in Eq. (2.3) and then use these estimates
to recover the estimates of resource shares for members of Iranian families by size.
Thereafter, I report the effect of family gender composition on each individual member’s
resource share. Finally, I report the estimates of the effect of demographics on overall
resource shares. Table 2.2 shows the coefficient estimates in Eq. (2.4) and their
standard errors for families in rural and urban areas, separately. The left panel reports
the parameters of equations for rural areas. The parameters of the quadratic equation in
family size for women, reported in the top panel, are 0.59, −0.25, and 0.049,
respectively. These parameters suggest that 59.2% of resources in childless families are
devoted to women. The coefficients of −0.25 and 0.049 for s and s2 suggest that a
woman’s share decreases sharply as family size grows. The estimate of γm
, which is the
effect of the proportion of girls in the family on the woman’s resource share, is 0.098,
which suggests that as the proportion of girls in the family increases, more resources are
devoted to the mother.
The second panel from the top in Table 2.2 reports the parameters of the
children’s equation. As mentioned in Section 3, this equation does not contain a
constant, as children’s resource share in childless families is zero. The coefficients of
family size and its square in the children’s equation are 0.182 and −0.01, respectively.
This result suggests that in one-child families, 17.1% of resources are devoted to that
child. The small coefficient of the squared term suggests that the share of each child
does not decrease considerably with the size of the family. The estimate of γc is −0.111,
supporting the existence of gender bias in resource allocation among children in rural
areas. In other words, the presence of girls in a family diverts resources away from the
children.
The bottom panel in Table 2.2 reports the resulting coefficients of resource share
for the father’s equation. The coefficients of the quadratic equation in s are 40.8, 0.07,
and −0.038, suggesting that households assign 40.8% of resources to the man in
childless families. In contrast to women, men retain or lose their resources smoothly as
family size grows.
30
Table 2.2: Estimates of Parameters
Rural Areas Urban Areas
Coefficient S.E. Coefficient S.E.
mothers
α0m
0.592*** 0.053 0.455*** 0.045
α1m
-0.252*** 0.066 -0.133** 0.064
α2m
0.049*** 0.018 0.027 0.019
γm
0.098*** 0.037 0.043 0.030
children
α1c
0.182*** 0.038 0.239*** 0.029
α2c
-0.011 0.017 -0.037*** 0.012
γc -0.111*** 0.037 0.001 0.032
fathers
α0f
0.408*** 0.053 0.545*** 0.045
α1f
0.070 0.071 -0.105 0.065
α2f
-0.038* 0.022 0.011 0.020
N 3343 4153
*: significant at 10%, **: significant at 5%, ***: significant at 1%.
The right panel in Table 2.2 reports similar results for families in urban areas.
The parameters of the quadratic expression in family size for women are 0.455, −0.133,
and 0.027, respectively. These estimates state that among childless families in urban
areas, the resource share devoted to women is 45.5%, which is smaller than the
estimate of women’s resource share among childless couples in rural areas. In addition,
it states that this resource share decreases sharply as families increase in size. The
decrease in women’s resource share by family size, however, is smaller than that seen
in rural areas. The estimate of γm
is roughly 0.043 and its sign is similar to that of the
rural areas analysis. Finally, the insignificant estimate of γm
means that children’s
gender composition does not influence the resources allocated to women in urban
areas.
31
The second panel from the top reports the estimates of the quadratic function of
family size for children. The estimates of 0.239 and −0.037 suggest that resource share
devoted to children is 20.2%, which is larger than the estimated share for the child in
one-child families in rural areas. The larger coefficient of the quadratic term relative to
that of rural areas suggests that the share of each child in two- and three-child families
decreases faster in urban areas.
The last panel reports the coefficients of the quadratic function in s for the man
as 0.545, −0.105, and 0.011. These estimates suggest that men’s resource share in
urban childless families is 54.5%, which is considerably larger than that in rural areas.
However, men’s resource share in urban areas declines at a faster rate as the number of
children rise.
2.4.1. Estimates of Resource Shares
Table 2.3 reports the estimates of resource shares by family size for families in
rural and urban areas, which derive from the coefficient estimates presented in Table
2.2. These estimates report the resource share for a reference family (see Section 4). All
the estimates in Table 2.3 are significant at the 1% significance level.
The top panel reports the estimates of resource shares for childless couples,
showing that 40.8% (54.5%) of family resources in rural (urban) areas are devoted to
men. The next panel (one-child families) highlights that the resource shares allocated to
men, women, and children in rural and urban areas are 43.9%, 39%, and 17.1% and
45.1%, 34.8%, and 20.1%, respectively. These results are important in two aspects.
First, unlike childless families, the estimated resource shares are similar between urban
and rural areas, except that families in urban areas allocate more resources toward their
children. Second, the resources allocated to children are mostly diverted from their
mother’s share because in one-child families, women’s resource share declines
considerably relative to that of the women in childless families. The man−woman gap in
resource shares is 4.9% in rural areas and 10.3% in urban areas, both in favor of men.
The next panel for two-child families reports that the man, woman, and children
in rural areas receive 39.4%, 28.7%, and 32% of family resources, respectively. The
corresponding estimates for urban areas are similar (37.7%, 29.5%, and 32.8%).
Despite a gap of 3% in the child’s resource share between one-child families in rural and
32
urban areas, this gap decreases to only 0.4% for each child in two-child families. The
gap between the resource share of men and women is 10.7% in rural areas compared
with 8.2% in urban areas.
Table 2.3: Estimates of Resource Shares by Household Size
Rural Areas Urban Areas
Coef. S.E. Coef. S.E.
childless families
Man 0.408 0.053 0.545 0.045
Woman 0.592 0.053 0.455 0.045
one-child families
Man 0.439 0.029 0.451 0.028
Woman 0.390 0.027 0.348 0.027
Children 0.171 0.025 0.201 0.022
two-child families
Man 0.394 0.043 0.377 0.046
Woman 0.287 0.039 0.295 0.044
Children 0.320 0.037 0.328 0.040
three-child families
Man 0.272 0.076 0.325 0.082
Woman 0.282 0.062 0.295 0.077
Children 0.446 0.079 0.381 0.075
N 3343 4153
All estimated coefficients are significant at the 1% significance level.
The last panel reports the resource allocation among three-child families.
Resources allocated to men, women, and children in rural areas are 27.2%, 28.2%, and
44.6%, and in urban areas are 32.5%, 29.5%, and 38.1%, respectively. The first
important aspect of these results is that the man−woman gap in resource share shrinks
considerably compared with families that have one or two children. Second, the
resources allocated to each child decline to 14.8% in rural areas and 12.7% in urban
areas. This finding suggests that consumption among Iranian households follows
economies of scale, which distinguishes the models underlying this study from the
literature that follows Working’s (1943) Engel curve method.
33
Two points are worth emphasizing. First, women’s resource share is larger than
the estimates of previous studies of developing countries. For example, Dunbar et al.
(2013) estimate women’s resource share to be about 37%, 22%, and 17% for Malawian
families with one, two, and three children, respectively. It is thus surprising to see that
among childless families in Iranian rural areas, women’s resource share exceeds that of
men’s. Two potential explanations worth mentioning in this regard: First, in rural areas
marriage is sometimes considered as a strategic bind between the groom’s and the
bride’s families. In this regard, the bride, in the husband’s family, is treated with extra
respect as a sign of respect for her families. The respect of husband family for the bride
in this case results in higher resource share for her. An alternative explanation is that the
results might be a consequence of sample restriction. In rural areas it is common for
newly married couples to live with groom’s family. The sample, however, excludes the
couples who live in an extended family; therefore, the remaining couples are those with
higher autonomy of women. It is because woman’s autonomy can enforce a separate
residence right after the marriage. This autonomy often comes from woman’s family
reputation relative to that of man.
The second point regarding the results is that as families grow parents’ resource
share decreases to compensate for the increase in children’s resource share. However,
the decline in women’s resource share by family size is more pronounced than that in
men’s resource share. This result differs from the finding of Dunbar et al. (2013) that
men’s resource share among Malawian families is almost unaffected by family size. For
example, these authors estimate that men’s resource share is about 46% in three-child
families. Therefore, women’s and children’s resource shares decline in response to a
growing family.
Various economic and cultural factors influence resource share in families (Das
Gupta et al., 2003), including occupation. In this study, I categorize families by using a
variable (farmer) that indicates whether either parent earns any income from a farming
job, including gardening, ranching, and other related tasks. Table 2.4 reports the
estimates of the resource shares for urban and rural areas by farmer indicator.
34
Table 2.4: Estimates of Resource Shares by Household Size and Farmer Status
rural/
nonfarmer
rural/
farmer
urban/
nonfarmer
urban/
farmer
Coef. S.E. Coef. S.E. Coef. S.E. Coef. S.E.
childless families
Father 0.331 0.072 0.449 0.077 0.537 0.045 0.680 0.136
Mother 0.669 0.072 0.551 0.077 0.463 0.045 0.320 0.136
one-child families
Father 0.405 0.043 0.495 0.041 0.448 0.029 0.499 0.124
Mother 0.468 0.040 0.354 0.040 0.356 0.028 0.309 0.112
Children 0.127 0.037 0.151 0.026 0.197 0.021 0.192 0.118
two-child families
Father 0.385 0.059 0.430 0.060 0.384 0.047 0.409 0.181
Mother 0.314 0.055 0.257 0.056 0.282 0.045 0.301 0.136
Children 0.301 0.051 0.313 0.049 0.335 0.040 0.289 0.181
three-child families
Father 0.269 0.104 0.254 0.096 0.345 0.081 0.411 0.321
Mother 0.208 0.089 0.260 0.085 0.241 0.081 0.297 0.161
Children 0.523 0.092 0.486 0.100 0.414 0.079 0.292 0.317
N 1870 1473 3913 240
An important feature of this table is that the farmer–urban sample consists of only
240 observations. The small size of this group results in insignificant estimates of
resource share for this sample. Table 2.4 highlights the fact that relative to nonfarmers,
farmer families in both rural and urban areas direct more resources toward the man and
fewer resources toward the woman compared with nonfarmer families. This finding
suggests that women in Iranian families that work in agriculture have less autonomy than
women in nonagricultural sectors. The resources allocated to children do not differ by
farmer indicator in rural areas, except in one-child families in which families devote fewer
resources to the child. In urban areas, however, farmer families allocate a smaller
resource share to children relative to nonfarmer families.
The fact that a large proportion of families in rural areas are involved in
agricultural production may create some concerns regarding the validity of separability of
35
production and consumption. However, the collective household model of BCL and DLP
does not require the separability of production and consumption to hold for all the goods
and services that an agricultural family consumes. It only requires the separability
assumption to hold for the private assignable good. That is, I assume that families
purchase their clothing and footwear. Given the fact that Iranian families in rural areas
are mostly involved in farming and carpet production, this assumption of the model is
consistent with Iranian context.
2.4.2. Children’sGender Composition
As stated in Section 4, the results presented thus far are estimates of resource
share for a reference family whose children are boys. In this section, I thus investigate
whether and to what extent the presence of girls in the family changes the resources
allocated to each family member. Table 2.5 presents the results separately for families
with one to three children. The top panel shows the effect of female children on resource
allocation among one-child families, highlighting that one-girl families in rural areas
allocate 3.8 percentage points more resources to women and 1.9 percentage points
fewer resources to children compared with one-boy families. The effect on fathers’
resource share is negative and insignificant.
A similar pattern exists among two and three-child families in rural areas. Two-
child families with two daughters devote 2.8 percentage points more resources to
women and 3.5 percentage points fewer resources to children relative to families with
two boys. Among three-child families, in families with three daughters, 2.8 (4.9)
percentage points more (fewer) resources are devoted to women (children) compared
with the reference family. Again, no significant effect of gender composition on men’s
resource share is evident.
The effects of children’s gender composition on intra-household resource
allocation among Iranian families in rural areas are thus threefold. First, 1.6–1.9
percentage points fewer resources are devoted to each girl relative to each boy. To
contextualize the extent of this effect, recall that the resource share of boys is 14.8–
36
17.1% depending on family size4. Second, the presence of girls in the family increases
women’s resource share by between 2.8 and 3.8 percentage points depending on family
size. Third, we observe no significant effect of family gender composition on men’s
resource share in rural areas. On the contrary, gender composition plays no significant
role in intra-household resource allocation in urban areas.
Table 2.5: The Effect of Female Children on Family Members’ Resource Shares
Rural Areas Urban Areas
Coef. S.E. Coef. S.E.
one-child families
Man -0.019 0.013 -0.015 0.010
Woman 0.038*** 0.014 0.015 0.010
Children -0.019*** 0.007 0.000 0.006
two-child families
Man 0.007 0.013 -0.013 0.011
Woman 0.028*** 0.010 0.013 0.009
Children -0.035*** 0.012 0.000 0.011
three-child families
Man 0.022 0.018 -0.013 0.013
Woman 0.028** 0.011 0.013 0.009
Children -0.049*** 0.018 0.000 0.012
N 3343 4153
**: significant at 5%, ***: significant at 1%.
The fact that presence of boys in a family is associated with lower income share
for the mother supports a competition scenario as discussed earlier. It shows that fathers
in rural areas may sacrifice their own resource shares for their male children. This result
is similar to that of Rose (1999) and DLP. DLP found that in all-girl families, the
combined share of children is about 6 percent lower than that of all-boy families and that
the extra resource share almost fully diverts to the women.
__________________________________ 4 These results follow from assuming a linear effect by modeling the effect of family gender composition using the proportion of girls.
37
The fact that such gender bias is evident only in rural areas reflects the
differences in economic and cultural factors. For example, gender differences in the
family’s contribution to the economy and old-age support for parents are among the
strongest economic motives for allocating resources in favor of boys. Allowing for
production in the model provides an alternative explanation to the results. Given the
possibility of working on a farm in rural areas, the difference in resource allocation
between boys and girls in urban vs. rural areas might reflect the difference between boys
and girls in their productivity in the two environments. In rural areas, boys may be more
productive than girls because they are more likely to work in farms and they may be
more productive in physical activities in the agricultural sector. Therefore, it may be
efficient for households to allocate higher resources to boys than girls. However, in
urban areas boys and girls similarly play more similar roles in the household. Therefore,
it may be efficient for households to allocate resources equally between girls and boys.
The view that boys and girls have different production function in rural areas and similar
production function in urban areas is supported by theoretical and empirical literature.
(Michael 1974; Polachek and Polachek 1989; Gugl and Welling 2012).
Similarly, the kinship system and construction of gender in society are among the
cultural norms that influence the roles of family members (Das Gupta et al., 2003). The
distinction between the driving forces behind the gender bias in rural areas lies in the
scope of their influence as well as policies that can alleviate such behavior. Contrary to
economic forces, cultural forces are wider in influence and usually require more time to
abate.
While distinguishing between the influence of economic and cultural factors is
challenging, one simple test is to produce the estimates by farmer indicator similar to
those in Section 5.1. The estimates of the influence of gender composition on resource
share by farmer status in Table 2.6 report the effects of female children on intra-
household resource allocation for four subsamples: rural–nonfarmer, rural–farmer,
urban–nonfarmer, and urban–farmer. These results show that the role of gender
composition in the nonfarmer sample in rural areas is considerably smaller than that in
the farmer subsample.
38
Table 2.6: The Effect of Female Children on Family Members’ Resource Shares
rural/
nonfarmer
rural/
farmer
urban/
nonfarmer
Farmer/ urban
Coef. S.E. Coef. S.E. Coef. S.E. Coef. S.E.
one-child families
Father -0.014 0.019 -0.037* 0.021 -0.019* 0.011 0.022 0.039
Mother 0.024 0.019 0.057*** 0.022 0.020* 0.011 -0.036 0.040
Children -0.010* 0.006 -0.020** 0.009 -0.000 0.006 0.014 0.028
two-child families
Father 0.007 0.017 -0.001 0.019 -0.015 0.011 0.014 0.046
Mother 0.016 0.013 0.041*** 0.015 0.016* 0.009 -0.035 0.041
Children -0.023 0.014 -0.041** 0.017 -0.001 0.011 0.022 0.042
three-child families
Father 0.029 0.025 0.021 0.029 -0.013 0.014 0.013 0.051
Mother 0.011 0.009 0.042** 0.018 0.013 0.008 -0.035 0.042
Children -0.040 0.025 -0.063** 0.027 -0.001 0.013 0.022 0.046
N 1870 1473 3913 240
*: significant at 10%, **: significant at 5%, ***: significant at 1%.
In nonfarmer families, the effect is significant only for children’s resource share in
one-child families, at −1 percentage point. Among farmer families, while the effect on
men in the farmer sample is still insignificant, except for one-child families, the effect on
women and children is more pronounced than the estimates presented in Table 2.5 for
aggregate rural areas. All-girl farmer families with one to three children in rural areas
devote 5.7, 4.1, and 4.2 percentage points more resources to women, respectively, each
of which is larger than the corresponding effect in aggregate rural areas (3.8, 2.8, and
2.8 percentage points). The resource share of children in all-girl farmer families with one
to three children in rural areas is 2, 4.1, and 6.3 percentage points smaller than that in
all-boy families, respectively. These effects are larger than those reported for all rural
areas (1.9, 3.5, and 4.5 percentage points).
Regarding the effect of gender composition on intra-household resource
allocation by farmer status for urban areas, the results presented in Table 2.6 show no
39
considerable difference between the nonfarmer sample and previous results on
aggregate urban areas5. These estimates support the important role of economic factors
in the presence of gender bias in intra-household resource allocation in rural areas.
Because the majority of families in Iranian rural areas are involved in agricultural
employment, boys are considered as a source of greater economic contribution to the
family relative to girls. Furthermore, family structure in rural areas is such that girls leave
the family after marriage and only boys provide for their parents in old age. These are
the principal reasons behind the fact that parents in rural areas treat boys more
favorably than girls in terms of intra-household resource allocation. By contrast, in urban
areas, the family system is more flexible and parents’ jobs are usually independent of
family members’ contributions. In addition, the more developed social infrastructures in
urban areas suggest a channel from which parents obtain old-age support. For example,
the development of pension funds plays a similar role to that of old-age support in rural
areas.
The fact that there are larger differences between boys’ and girls’ resource
allocation in rural farmer families than in non-farmer families is consistent with the idea
that there may be productivity differences between boys and girls in rural agriculture that
result in the allocation of more resources to boys.
Other studies have reported on the difference between rural and urban areas in
terms of son preference. For example, Long (1989) finds that the sex ratio in Shanghai
(a major city) has become fairly normal at about 1.05 in sharp contrast to the inflated sex
ratios in rural areas and smaller cities, which reflect the higher mortality rate of girls and
is associated with son preference. In the same vein, Rose (1999) reports the difference
in parents’ son preference in India by state, reporting the sex ratio at birth to be as low
as 0.99 and 1.06 in Assam and Kerala but higher than 1.4 in West Bengal, Punjab, and
Uttar Pradesh.
Iranian families devote more resources to sons than to daughters for one of three
reasons. First, Iranian culture is strongly patriarchal (Moghadam 1992; Rasul-Ronning
2013). Fathers typically make the final decision on important family matters. In some
instances, grandfathers even wield power over large extended families. In such a
cultural environment, it is natural for parents to value the son as the next generation’s
father. In addition, children in Iran get their names from their father’s side. Therefore, son
__________________________________ 5 As mentioned in Section 5.1, the farmer sample in urban areas includes only 240 observations, which results in imprecise estimated coefficients.
40
preference can be viewed as a way of retaining the family name. In this context, the first-
born has a more important role in the family and sons are even more valuable if born
first. Indeed, in more traditional areas, a mother is more respected if she gives birth to a
boy first (Moghaddam, 1992).
Second, a woman usually leaves the family after marriage and moves in with the
husband (or husband’s family). The son, however, typically supports the family,
especially as his parents age. In multi-son families, this role is assigned to the oldest
son. Third, Iranian women have low labour market participation (Mehryar et al., 2002;
Majbouri, 2010). Over the past three decades, female labour force participation has
been below 20%, suggesting that men are the main economic providers for the majority
of families. Although some of the abovementioned mechanisms, including the kinship
system, are evolving, driven by urbanization and economic development, it is reasonable
to expect that cultural effects will continue to influence the behavior of people in Iran.
Given the infrequency of reported purchases in HIES, total expenditure can be
subject to measurement error. Dunbar et al. (2013) address this via a generalized
Method of Moments estimator using household income as an instrument for total
expenditure. However, they found that their results regarding the intrahousehold
resource allocation and gender bias are invariant to instrumenting for total expenditure.
I address this concern by instrumenting for total expenditure in a control function
framework (Heckman and Robb, 1985). The instruments are several measures of
wealth, including family income from different sources (wage and salary, self-
employment, and non-labour income), home ownership and car ownership.
I apply the control function by regressing the logarithm of total expenditure on
instruments and demographics. Then, I use the residual from the first stage as a linear
shifter in preference in demand equations for each family member. I should note that a
Generalized Method of Moments approach is preferred as it is consistent even when the
system of equation cannot be specified as a triangular form. For further discussion
regarding the control functions and systems of equations please refer to Blundell et. al
(2013).
Table 2.7 shows the effect of family gender composition on resource shares. This
table supports two main findings of this study. First, fewer shares of family resources are
devoted to female children in rural areas. Second, the lower share of girls in rural areas
is associated with higher shares for their mother. Third, family gender composition does
not play any role in intra-household resource allocation in urban areas.
41
Table 2.7: The Effect of Female Children on Family Members’ Resource Shares
Rural Areas Urban Areas
Coef. S.E. Coef. S.E.
one-child families
Man -0.017 0.016 -0.016 0.011
Woman 0.035*** 0.016 0.016 0.011
Children -0.018*** 0.007 -0.001 0.007
two-child families
Man 0.012 0.016 -0.013 0.012
Woman 0.024*** 0.011 0.014 0.009
Children -0.035*** 0.014 -0.001 0.011
three-child families
Man 0.026 0.023 -0.014 0.014
Woman 0.027** 0.013 0.015 0.011
Children -0.052*** 0.021 -0.001 0.012
N 3343 4153
**: significant at 5%, ***: significant at 1%.
2.4.3. Effect of Demographics
The effects of demographics on resource share reported in Table 2.8 show that
paternal education in cities diverts resources from mothers to fathers. Each extra year of
the husband’s education increases his resource share by about 0.67 percentage points,
all of which is diverted from the mother’s resource share. Maternal education in rural
areas diverts the father’s resource share to children (0.4 percentage points). In urban
areas, it diverts resources from fathers (0.65 percentage points) to mothers (0.61). For
example, if a woman in an urban area moves from the median education level (grade 5)
to a high school diploma level (grade 12), her resource share increases by 4.2
percentage points, her husband’s resource share decreases by 4.5 percentage points,
and her children’s resource share remains almost unchanged. This result is consistent
with a bargaining mechanism among couples in the allocation of resources, in which
education provides the educated party with higher bargaining power. Education thus
acts as a distributional factor that influences resource share. Another interesting finding
regarding maternal education in rural areas is that it diverts resources from fathers to
children and results in a significant positive impact on children’s resource share. In this
42
regard, Dunbar et al. (2013) find that maternal education diverts resources from fathers
to mothers (60%) and to children (40%). However, paternal education in the Dunbar et
al. (2013) study does not have a substantial effect on family members’ resource shares.
Table 2.8: The Effect of Demographics on Intra-household Resource Allocation
Rural Areas Urban Areas
Coef. S.E. Coef. S.E.
man education
Man 0.0012 0.0036 0.0067** 0.0032
Woman -0.0005 0.0035 -0.0064** 0.0031
Children -0.0007 0.0011 -0.0003 0.0007
woman education
Man -0.0042 0.0042 -0.0065* 0.0034
Woman 0.0004 0.0041 0.0061* 0.0034
Children 0.0038** 0.0013 0.0004 0.0007
man age
Man 0.0003 0.0025 -0.0022 0.0024
Woman -0.0002 0.0024 0.0021 0.0024
Children -0.0001 0.0006 0.0001 0.0005
woman age
Man 0.0038 0.0025 0.0037 0.0025
Woman -0.0043* 0.0025 -0.0036 0.0025
Children 0.0005 0.0007 -0.0000 0.0005
children’s average age
Man -0.0097* 0.0053 -0.0031 0.0055
Woman 0.0098* 0.0052 0.0034 0.0054
Children -0.0001 0.0008 -0.0003 0.0006
children age difference
Man -0.0010 0.0091 0.0070 0.0118
Woman -0.0142* 0.0077 -0.0058 0.0089
Children 0.0152** 0.0075 -0.0012 0.0115
N 3343 4153
*: significant at 10%, **: significant at 5%.
43
The substantial positive effect of maternal education on children’s resources is
important for formulating policies on reducing child poverty. Previous studies in
developing countries also support that improvements in women’s status can contribute
to their children’s well-being and outcomes (Hill and King, 1995).
A woman’s age in rural areas has a negative impact on her resource share. As
women age, they lose their productive power and, consequently, their bargaining power
in the family. For each extra year of age, women’s resources are reduced by 0.43
percentage points. For instance, a 35-year-old woman in a rural area loses 2.1
percentage points of her resource share as she approaches 40 years. This finding
contrasts with that of Das Gupta et al. (2003), who report that women’s power
dramatically rises over their lifespan as power and autonomy shift from fathers to
mothers in older age.
Older children in rural areas divert resources toward their mothers. For example,
one extra year of the average age of children diverts about 1 percentage point of
resources from fathers to mothers. Raising the average age of children from five to 10
years in rural areas also diverts resources from fathers (5 percentage points) to mothers.
This finding, that older children divert resources from fathers to mothers, is consistent
with the results of Dunbar et al. (2013) from Malawian rural areas. The higher diversity of
children’s ages diverts resources mostly from mothers to children in Iranian rural areas.
One extra year of children’s age diversity increases their resource share by 1.5
percentage points, which is diverted mostly from their mother’s resources. In rural
Malawi, Dunbar et al. (2013) find that a higher variance in children’s ages diverts
resources from fathers to mothers.
A comparison of the effect of covariates in rural and urban areas suggests that in
urban areas, the most influential variable in intra-household resource allocation is
parents’ education. In rural areas, however, education plays less of a role than children
and women’s age. This finding supports the notion that in rural areas, inter-family
bargaining power is governed by traditional social norms, which values family members
according to their age and productivity. On the contrary, educational attainment has a
small role in bargaining power in rural areas compared with urban areas, in which
education plays a more important role than age and physical productivity.
44
2.5. Conclusions
This study estimated family members’ resource shares and investigated gender
bias in intra-household resource allocation among Iranian families. Methodologically, it
extended the approach taken by Dunbar et al. (2013) to allow for resource share to
depend on family gender composition as well as on total expenditure and family size.
The estimates of resource share presented herein suggest that among families
with one or two children, there is a gap between the resources allocated to men and
women in favor of men. However, this gap shrinks in three-child families. In childless
families in urban areas, the gap is small, while women’s resource share outweighs that
of men in rural areas. These results suggest that women, especially among families with
one and two children, trade off more between their own and their children’s
consumption. Children’s resource share increases as family size grows, from 17.1%
(20.1%) in one-child families to 44.6% (38.1%) in three-child families in rural (urban)
areas.
In addition, the study found that intra-household resource allocation in rural areas
is influenced by gender composition. Parents assign 1.6–1.9 percentage points more
resources toward their sons relative to their daughters, an effect more pronounced
among farmer families. Such strong gender bias among farmer families in rural areas
suggests that economic factors are more influential in the presence of parents’
discriminatory behavior toward their children. Male children in farmer families are a
source of greater contribution relative to female children and, therefore, families treat
boys relatively more favorably.
Similarly, this study found that women in all-boy families in rural areas are
assigned 2.8–3.6 percentage points fewer resources than those in all-girl families. As
before, these effects are higher among farmer families. The fact that the presence of
male children has a negative effect on their mother’s resource share suggests either that
the reward argument is not plausible for explaining the effect of gender composition on
women’s resource share or that the effect is influential in the short run (i.e., it does not
change women’s lifelong autonomy). In contrast to the rural sample, the study found no
significant role of gender composition on intra-household resource allocation in urban
areas.
Different effect of gender composition on intra-household resource allocation
between rural and urban areas might suggest that family member’s production functions
45
are different in rural and urban areas. Likewise, the difference between farmer and
nonfarmer families in production function might be an explanation for more pronounced
effect of gender composition on intra-household resource allocation in farmer families.
Boys in rural areas considered to be more productive than girls in farm production and
later as old age support for parents. Therefore, an efficient resource allocation will
allocate more resources toward boys. In such environments, the extra resource share of
boys comes at the expense of mothers as they are considered less productive than
fathers. In urban areas, however, there is no difference in production function of boys
and girls. Consequently, we do not observe any evidence of gender bias in intra-
household resource allocation in urban areas.
Finally, the fact that maternal education has a substantial effect on children’s
share has important policy implications for the alleviation of child poverty.
46
Chapter 3. Private Schools and Student Achievement 6
Critics frequently argue that public schools are hamstrung by a variety of
constraints that undermine their effectiveness, including enrolment rules that restrict
competition, top-down management structures that limit their discretion to innovate, and
strong union rules that prevent them from disciplining or dismissing poor teachers.
Private schools, in contrast, are forced to compete in order to attract tuition-paying
students, can dismiss teachers who underperform, and may apply selective enrolment
rules in order to generate complementarities among students. With greater access to
resources, private schools may be able to hire better or more experienced teachers, and
offer smaller classes and programs that are specially designed to serve their students.
Given these advantages, economists would expect private schools to provide
higher quality education than their public counterparts. Surprisingly, however, the
empirical evidence consistently finds that private schools in advanced economies do not
do a better job of teaching core reading and numeracy skills than public schools, and
may in fact do worse on average7. Highly credible random assignment studies of the
effects of tuition fee voucher programs in the United States find little if any effect of
__________________________________ 6 This chapter is a joint work with Dr. Jane Friesen and Dr. Simon Woodcock.
7 Evidence of the relative performance of private versus public schools in developing economies is more favorable. See, for example, estimates of private school effects for Columbia (Angrist 2002, 2006), India (Muralidharan and Sundararaman 2015; Singh 2015), Pakistan (Andrabi et al. 2011; Alderman et al. 2001).
47
private schools on students’ test scores, with the possible exception of African-American
students (see Epple et al. (2015a) for a review). A second literature that uses
instruments based on geography and/or religiosity to address selection into private
Catholic high schools finds very little evidence of any effect on test scores, except
among urban minorities (e.g. Figlio and Stone 1997; Grogger and Neal 2000).8 Along
with others, Altonji et al. (2005) question the validity of these instruments. They instead
control for as many observable student characteristics as possible in a value-added
model of achievement, and then generate bounds on the true causal effect under
plausible assumptions about the relationship between selection on observable and
unobservable student characteristics. Like the previous IV studies, they find no effect of
private Catholic high schools on test scores in the U.S. Applying Altonji et al.’s method,
Elder and Jepsen (2014) find negative effects of private Catholic primary schools on
numeracy scores in the U.S., and Nghiem et al. (2015) find negative effects of private
Catholic primary schools on both reading and literacy scores in Australia. Earlier studies
that use the ECLS-K data to study the effect of private Catholic schools on primary test
scores find small negative and insignificant results (Carbonaro 2006; Lubienski et al.
2008; Reardon et al. 2009). Jepsen (2003) finds small positive effects of Catholic
primary schools on high school test scores in data from the Prospects project.
Recent evidence of the effects of secular private schools is more limited but
somewhat more promising. Figlio and Stone (1997) find positive effects of private
secular high schools in the U.S. on mathematics and science scores. Nghiem et al.
(2015) find generally positive but statistically insignificant effects on numeracy and
literacy skills in Australian private secular primary schools. Lefebvre et al. (2011) control
for individual heterogeneity via student fixed effects rather than lagged test scores; they
find positive effects on private secular high school numeracy scores in Quebec, Canada.
__________________________________ 8 A number of papers find that private schools do, however, confer an advantage with respect to high school completion (e.g. Evans and Schwab 1995; Grogger and Neal 2000; Neal 1997).
48
Our paper advances this literature by providing the first evidence of the effects of
private schools on student achievement based on longitudinal population data. This data
affords several advantages over previous studies of private school effects on
achievement. First, it provides a much larger sample,9 allowing us to contribute new and
precise estimates to the literature on the effectiveness of Catholic and secular private
schools, and to extend this literature to include other types of faith schools. Second,
because it includes both a substantial number of observations concentrated within fairly
small geographic areas and detailed information about students’ residential locations, we
can control for neighborhood characteristics or neighborhood fixed effects in our
specifications. Doing so allows us to control for an important potential source of
unobserved heterogeneity in family background and student ability that is correlated with
neighborhood choice. Moreover, these neighborhood controls ensure that private school
effects are identified either from comparisons of schools attended by students who live in
similar types of neighborhoods (when we include neighborhood characteristics), or who
live in the same neighborhoods (when we include neighborhood fixed effects). This
feature is important: some previous studies report positive effects of U.S. Catholic
schools when they restrict their attention to urban populations, although sample sizes
become very small. A common explanation offered for this pattern of results is that the
quality of public schools is lower in urban areas than elsewhere, so that local private
schools can be superior to these schools while not outperforming public schools in
general (see Martinez-Mora, 2006, for a theoretical treatment of this issue). By
identifying private school effects from comparisons between schools attended by
students who live in nearby or similar neighborhoods, our approach avoids this
potentially confounding factor. A third advantage of using multiple cohorts of population
data is that it provides us with a sufficiently large number of observations per school to
__________________________________ 9 Specifically, our data include more than seven times the number of students enrolled in private Catholic schools compared to the data used by Nghiem et al. (2015) and Elder and Jepsen (2014), and more than nine times the number of students enrolled in private secular schools compared to the data used by Nghiem et al. (2015) and Lefebvre et al. (2011). McKewan (2001) has a larger sample of private Catholic school students in his data from Chile, but only one test score per student. Other jurisdictions that provide researchers with access to population-based longitudinal student-level data typically do not include records for students enrolled in private schools (e.g. Florida, England, North Carolina and Texas).
49
estimate individual school effects, allowing us to characterize heterogeneity both within
and across school sectors. Finally, the density of our data within one contiguous
geographic area allows us to estimate a full set of school and student fixed effects using
methods developed by Abowd et al. (2002). We use these estimates to characterize the
entire distribution of student as well as school quality both within and across school
sectors.
Our student-level administrative data from fourteen school districts in British
Columbia, Canada encompasses the population of students enrolled in virtually all
private as well as public schools. Our longitudinal records follow five cohorts of grade 4
students (aged 8-9) for a total of four years, when most of these students are in grade 7.
British Columbia (B.C.) administers standardized tests in reading and numeracy to
students in grades 4 and 7. Along with information about test scores and basic
demographic characteristics, individual records include information about the school
attended and the student’s residential postal code in each year.
B.C.’s public education system has many standard features and offers
considerable choice within the public sector to a highly diverse student population. B.C.
students consistently rank very well on international comparisons of math, science and
reading achievement (Brochu et al. 2012). Almost 12% of B.C. primary school students
were enrolled in private schools in seventh grade in 2010/11 (British Columbia Ministry
of Education 2011), roughly comparable to national private school enrolment rates in the
United States (Rouse and Barrow 2009). Most private schools have been operating for
many years, and have received public funding since 1977. This universal voucher,
typically worth half of the basic allocation provided for each student attending a public
school, is available to both religious and secular private schools. In order to be eligible
for the voucher, private schools are required to hire qualified teachers, offer the
approved core curriculum and meet minimum requirements for instructional time, and
they are subject to provincial inspection and evaluation. They are fully autonomous with
respect to personnel decisions, pay, school calendar, and admissions criteria.
50
We estimate several different specifications of the education production function,
using both a lagged value-added model and a student fixed effects model, and conduct
a number of robustness checks based on sub-samples of the data. We obtain a number
of important results that are consistent across specifications and samples. We find that
private schools on average outperform public schools by at least 0.10 standard
deviations in both reading and numeracy, and these results cannot be accounted for by
reasonable assumptions about the degree selection on unobservable factors relative to
selection on observable student and neighborhood characteristics.
Unlike most previous studies, we find that Catholic private schools outperform
their public counterparts in both math and reading, with an effect size of at least 0.10
standard deviations. Non-Christian faith private schools perform even better, with an
effect size of at least 0.20 standard deviations. Other Christian (i.e. non-Catholic) private
schools, in contrast, do not outperform public schools; estimated effects are small and
statistically insignificant. We are also able to obtain precise and robust estimates of the
effects of a group of secular private schools that we refer to as “prep” schools; these are
schools that emphasize academic achievement. We find that prep private schools on
average perform very well – somewhat better than Catholic private schools and as well
as non-Christian faith private schools.
3.1. Related literature
Several related literatures yield results that can inform our expectations of private
school performance relative to public schools. Clark (2009) investigates the performance
of “grant maintained” public schools in England. These schools have full authority to hire
and fire teaching staff and can apply some criteria for selecting students, but they are
not permitted to charge tuition fees or use admissions tests. He finds that converting to
grant-maintained status increased achievement gains by 0.25 standard deviations.
Changes in student characteristics can account for less than half of this improvement;
the remaining gains appear to be associated with turnover in teaching staff and an
increase in the number of teachers.
51
In the U.S., the charter school movement has grown as an alternative to regular
public schools. Charter schools are granted the freedom to innovate with respect to
curriculum and pedagogy. Compared to regular public schools, they are less likely to be
unionized and exhibit higher rates of teacher turnover (Epple et al. 2015a). Unlike private
schools, charter schools cannot charge tuition and must use lotteries to admit students if
they are oversubscribed. Lottery-based studies of oversubscribed charter schools tend
to find fairly large positive effects. Estimates of a broader set of charter school effects,
including those that are not oversubscribed, find no overall effect, and sometimes
negative effects in the early years of newly established charter schools (see Epple et al.
(2015b) for a review).
3.2. Institutional Context
3.2.1. Public school choice and funding
The majority of public primary schools in British Columbia offer Kindergarten
through grade 7, with high schools offering grades 8 through 12. There are many
exceptions, however, with some schools offering Kindergarten through grades 3, 4, 5 or
6, some middle schools that begin in grade 6 and some “junior” high schools that begin
in grade 7.
Students in B.C. are guaranteed access to a “catchment” public school based on
their residential address. They may also choose to enroll in a regular public school other
than their catchment area school. Before July 2002, the provincial education authority
(the Ministry of Education) mandated that out-of-catchment enrollment in a regular (non-
magnet) public school required permission of the principals of both the catchment area
school and the preferred school. Since July 2002, students have been free to enroll in
any public school in the province that has space and facilities available after students
who reside in the catchment area have enrolled. Transportation to non-catchment
52
schools is not provided. When catchment schools are over-subscribed, provincial
legislation requires that school boards give priority to students who reside within the
district. Boards may elect to give priority to siblings of children who are already enrolled.
Within these enrolment categories, principals of regular public schools have discretion
over which students to enroll.
Parents in B.C. may also choose to enroll their children in a public magnet
program. The most popular form of magnet program is French Immersion, which enrolls
about 10 percent of Kindergarten students in the province (BC Ministry of Education
2011). Entry into French Immersion programs is restricted to students entering
Kindergarten or grade 1, and space is often allocated by lottery.
The B.C. Ministry of Education provides operating and capital funding directly to
public districts. Operating funds are provided in proportion to total district enrolment, with
supplementary funding for each student who is Aboriginal, gifted or disabled, or who
qualifies for English as a Second Language (ESL)10 instruction. Public districts and
schools are not authorized to raise any additional revenue, and are required to offer the
provincial curriculum. Hiring, firing and remuneration of teachers is governed by strict
rules specified in a collective agreement between the Province and the powerful union
that represents B.C. teachers (the British Columbia Teachers Federation).
3.2.2. Private school choice and funding
Since 1977, British Columbia has provided public grants to private schools that
conform to provincial curriculum standards and meet various provincial administrative
requirements (B.C. Federation of Independent Schools11 Associations 2015). The value
of the grant varies with school operating costs: since 1989, schools whose operating
costs are no higher than in the public system (Group 1 schools) receive an amount per
__________________________________ 10
The ESL program was renamed to English Language Learning (ELL) in January 2012. 11
The “independent school” has been widely used in BC province to refer private schools. We use both terms interchangeably.
53
student equal to 50 percent of the value of the per student grant to public schools; those
with higher operating costs (Group 2 schools) receive 35 percent of the public school
grant (B.C. Ministry of Education 2005). The Ministry of Education does not limit the total
number of funded private school spaces, and private schools are not constrained in their
selection of students. The formula for supplementary funding for special education
students in private schools changed in 2005. Private schools had historically received
half as much per student special education funding as public schools; since 2005,
private schools have received the full value of the public school special education
supplement.
In order to be eligible for funding, private schools must hire qualified B.C.
teachers and offer the provincial curriculum. Private schools may charge tuition, apply
any admissions criteria that do not violate the Canadian Charter of Rights and Freedoms
or the provincial Human Rights Code, and can hire, fire and remunerate teachers subject
only to provincial labour standards legislation.
3.2.3. Testing and accountability
All public and provincially funded private schools in British Columbia are required
to administer standardized tests to students in grades 4 and 7 in reading and numeracy
each year. These scores do not contribute to students’ academic records and play no
role in grade completion, and there are no financial incentives for teachers or schools
related to student performance. The Ministry of Education began posting school-average
test scores on their website in 2001 (B.C. Ministry of Education 2001). The Fraser
Institute, an independent research and educational organization (Fraser Institute 2008),
began issuing annual “report cards” on B.C.’s elementary schools in June 2003 (Cowley
and Easton 2003). These reports include school scores and rankings based on test
scores. From the outset, the school report cards have received widespread media
coverage in the province’s print, radio and television media.
54
3.3. Data
Our estimates are based on extracts from two administrative databases collected
and maintained by the B.C. Ministry of Education. The first is an enrolment database that
records the school at which each student is enrolled on September 30 of each year. Our
extract includes five cohorts of grade 4 students who were enrolled in a public or private
school located within the geographic boundaries of the fourteen school districts in the
Lower Mainland of B.C.12 in 1999/2000 through 2003/2004, and follows them for the
following four years. Students remain in our data so long as they remain within the
provincial public or private school system in any of these districts. The individual records
include indicators for the language spoken in the student’s home (English, Chinese,
Punjabi, and other), whether the student self-identified as Aboriginal in any year,
whether the student was registered in ESL or special education (i.e. a gifted or disabled
program), whether the student was enrolled in French Immersion, whether the school is
public or private, and the student’s gender. In addition, the extract provides the student’s
residential postal code and unique student, school and district identifiers. We attach
average family income, proportion of immigrant families, and proportion of people with
different levels of education in the student’s Census neighborhood (enumeration area),
based on a postal code match. An enumeration area is the smallest geographic area for
which public-use Census data are produced, and typically comprises several hundred
households. A detailed description of our procedures for locating residential postal codes
within enumeration areas is provided in a data appendix.
The second database provides student-level data on participation and scores on
standardized tests administered in grades 4 and 7 for the 1999/2000-2006/2007 school
__________________________________ 12 The Lower Mainland consists of the city of Vancouver and its suburbs. It is geographically
isolated by the Canada/U.S. border to the south, rugged mountains to the east and north, and the Salish Sea to the west.
55
years. We merge students’ test scores with the enrolment database via the unique
student identifier provided in both files.
3.4. Methodology
3.4.1. The value-added model
The literature on specification issues related to education production functions is
highly developed (e.g. Todd and Wolpin 2003; Andrabi et al. 2011). The most commonly
used specification in the education literature is the lagged value added model, which
includes a lagged test score as a sufficient statistic for the entire history of inputs in a
model of current outcomes. We write our basic version of this model as:
(3.1)
where is an outcome measure observed when student i is in grade 7, is a
vector of student characteristics observed in grade 7, is an indicator that
student i attends a private school in grade 7, is a fixed effect associated with the
year that student i is in grade 7 and is a stochastic error. The parameter of interest,
captures the average contribution of private versus public schools to student test score
growth between grades 4 and 7.
The key identifying assumption is that unobserved factors affecting grade 7 test
scores are mean-independence with respect to enrolment in a private versus public
school, conditional on grade 4 test scores and other observable student characteristics:
(3.2)
56
This condition requires that there are no unobserved factors that affect test score
growth that are systematically related to the decision to attend a private versus public
school in grade 7, conditional on observable student and neighborhood characteristics.
A further threat to identification comes from the inclusion of the lagged dependent
variable in this model, since test scores are prone to measurement error. However,
despite their theoretical limitations, value-added models that include an adequate set of
controls have been shown to deliver estimates of the effects of various inputs with a
fairly small degree of bias in a range of contexts.13 We include a number of individual
student characteristics, including home language, Aboriginal identity and gender. We
use two alternative sets of controls based on the student’s residential neighborhood. In
one case we add mean neighborhood socioeconomic characteristics, including family
income, parents’ education and proportion of immigrants. In another case we include
neighborhood (postal code) fixed effects. Gibbons and Silva (2011) find that including
home postal code fixed effects as controls in their value-added model accounts for a
substantial amount of the estimated differences in quality between faith-based and
secular public schools in England.
We begin by estimating this basic model, which provides an estimate of the
difference between the average test score of students attending private versus public
schools. We then extend this model to include a full set of school fixed effects. The
condition for identification in this case can be stated analogously to (3.2), with the private
school indicator replaced by a full set of school fixed effects. In this case, identification
requires that there are no unobserved factors that affect test score growth that are
__________________________________ 13 Several recent studies have compared the results from value-added estimators to experimental
results from the same data set. Andrabi et al. (2011) find biases from measurement error and unobserved heterogeneity are offsetting in their study of private schools in Pakistan, such that the aggregate bias on the private school coefficient is not significant. Deming et al. (2014) find no significant differences between experimental estimates of school effects based on lottery data and estimates from a value-added model that controls for a previous test score. A number of studies find similar results when comparing experimental estimates to value-added estimates of teacher effects (Angrist et al. 2013; Kane and Staiger 2008; Kane et al. 2013).
57
systematically related to the decision to attend each private versus public school in
grade 7, conditional on observable student and neighborhood characteristics.
3.4.2. The student fixed effects model
We next estimate an alternative model of the education production function, in
which the coefficient on the lagged score, is constrained to be zero. This restriction is
supported by growing evidence that the effects of lagged inputs dissipate rapidly (see for
example Andrabi et al. 2011, Jacob et al. 2010, and Kane and Staiger 2008 in the
context of teacher effects). Since we observe test scores three years apart, any bias
introduced by this restriction is likely to be very small. We control for individual
heterogeneity by including student fixed effects. The student fixed effects specification
has been used widely in the school quality literature, including several recent papers on
the effects of charter schools (e.g. Imberman 2011). We write this model as:
g=4,7 (3.3)
where is a grade-by-year fixed effect.
The key identifying assumption in the student fixed effects model is that, among
the subset of students who switch between private and public schools between grades 4
and 7, unobserved factors that affect a student’s test score in a given grade are
uncorrelated with private school status in that grade, conditional on grade-level transitory
shocks, student fixed effects and observable time-varying student characteristics:
for g = 4, 7 (3.4)
58
At least two plausible scenarios could threaten the validity of this assumption.
First, students may be heterogeneous with respect to an unobserved effect in test score
growth between grades 4 and 7. The student fixed effects estimator will be biased if this
effect is correlated with patterns of student mobility. Suppose, for example, that students
who switch from a public school in grade 4 to a private school in grade 7 on average
have characteristics associated with higher rates of test score growth, all else equal,
than those who switch from private to public. In this case, the estimator will attribute the
higher rate of test score growth that is caused by this pattern of unobserved
heterogeneity to the effect of the private school.
Second, transitory shocks that cause students to change schools may be
correlated with unobserved factors that affect test scores. Suppose, for example, that
some students change schools at the end of grade 4 following a family break-up or job
loss, and this event also adversely affects student achievement in grade 4. If the
student’s grades recover by grade 7, the estimated quality of the grade 4 school will be
biased downwards relative to the grade 7 school. If students enrolled in private schools
in grade 4 experience this scenario with the same frequency and degree as those
enrolled in public schools in grade 4, the estimated difference between public and
private school quality will not be affected. If, for example, these types of shocks more
frequently result in student moving from private school to public schools, rather than vice
versa, the estimated quality of private schools will be biased downwards.
We address this second selection problem in two ways. First, we investigate the
sensitivity of our results to several sample restrictions designed to eliminate bias
associated with different patterns of correlation between mobility patterns and transitory
shocks to student achievement. By excluding students who changed schools
immediately after grade 4, we eliminate the threat of bias from shocks that precipitate a
move at the end of grade 4 and affect grade 4 test scores. By excluding students who
changed schools immediately after grade 6, we eliminate the threat of bias from shocks
that precipitate a move at the end of grade 6 and affect grade 7 test scores. Second, we
estimate our model for a sample of students who are required by the grade configuration
of schools to move between grades 4 and 7. The types of transitory shocks that lead to
59
the dynamic selections problems described above are likely to be less frequent among
compulsory movers than among students who elect to change schools.
Again, we begin by estimating this basic model, and then extend the model to
include a full set of school fixed effects. The condition for identification in this case can
be stated analogously to (3.4), with the private school indicators replaced by a full set of
school fixed effects for each grade. In this case, identification requires that, among the
subset of students who switch between private and public schools between grades 4
and 7, unobserved factors that affect a student’s test score in a given grade are
uncorrelated with the choice of school (rather than school type) in that grade, conditional
on grade-level transitory shocks, student fixed effects and observable time-varying
student characteristics. We estimate this two-way fixed effects model using a procedure
developed by Abowd et al. (2002).
3.4.3. Bounding the effects of selection on unobservables
None of these approaches fully addresses the threat from unobserved individual
fixed effects in test score growth. Following some of the recent literature (e.g. Nghiem et
al. 2015; Elder and Jepsen 2014), we generate bounds on some of main estimates
under assumptions about the relationship between selection on unobservable and
observable characteristics, using the procedure introduced by Altonji et al. (2005) and
extended by Oster (2015).
Defining the effects of observables as , we can write
the value-added model with a simple private school indicator as:
(3.5)
Call the estimate of obtained from this “long” version of the regression model ,
and the corresponding , .~R Now consider a “short” regression that includes only the
private school indicator:
60
( 3.6)
Call the estimate of obtained from this model , and the corresponding , .
Oster (2015) shows that
a -a
a0 -a= dRmax -R
R-R0 ( 3.7)
where is the degree of selection on unobservables relative to the selection on
observables and maxR is the value of we would obtain if we included all of the
observable and unobservable factors that affect the dependent variable. We can obtain
estimates of and and the corresponding values of R~
and 0R by estimating models
(3.6) and (3.7). Oster notes that the value of maxR will be less than one if there is
measurement error in the dependent variable. For an assumed value of maxR , we can
calculate the value of , the ratio of the degree of selection on unobservables to
selection on observables, under the null hypothesis that =0. Oster (2015) shows that
this method of calculating is equivalent to the method proposed by Altonji et al. (2005)
under the assumption that .1max R Both Oster (2015) and Altonji et al. (2005) argue
that it is unreasonable to expect the degree of selection on unobservables to exceed the
degree of selection on observables. A value of therefore would lead to the
conclusion that the estimated treatment effect is too large to be solely a consequence of
bias due to selection on unobservables. We use this method to calculate values of
using RR~
3.1max , as recommended by Oster (2015).
61
3.5. RESULTS
Our estimation sample consists of all students in the population of interest who
are observed in our data in grades 4, 5, 6 and 7. We restrict our attention to students
attending an English language14 public or private school that enrolls at least five students
in the relevant grade and year, and who have non-missing values for all relevant
variables.
3.5.1. Descriptive statistics
Table 3.1 presents selected school characteristics by school type. Our estimation
sample includes a total of 649 schools, of which 554 are public and 95 are private. Of
the private schools, 33 are Catholic, 39 are other Christian, 7 are other faith, and 15 are
secular prep schools. Three-quarters of private schools are in funding group 1, receiving
a per student subsidy equal to 50% of the public school subsidy; the remaining private
schools are in funding group 2, receiving a 35% subsidy. A small number of other (non-
prep) secular private schools offer Montessori or Waldorf programs, or specialized
education for students with special learning needs. Given their small number, typically
small size, and the diversity of programming, we exclude these schools from our
analysis. Almost all private schools in our sample offer both grades 4 and 7. In contrast,
101 public schools offer grade 4 but do not offer grade 7, and 35 schools offer grade 7
but not grade 4.
Table 3.2 presents descriptive statistics for the grade 7 students in our estimation
sample who have non-missing test scores. Private school students on average earn
higher test scores than public schools students, but this difference varies across private
__________________________________ 14 French language instruction is offered to English speaking students via public French
Immersion programs. Attrition from these programs is very high (see Shack 2015). Francophone students may choose to attend one of a small number of schools operated by the public francophone school board.
62
school types. Students enrolled in private prep schools excel; their average grade 7 test
scores are 0.86 and 0.93 standard deviations above the mean in reading and numeracy.
Students enrolled in faith private schools also score relatively well, averaging 0.43
standard deviations above the mean in reading and 0.54 standard deviation in numeracy
at private Catholic schools, 0.27 standard deviations in reading and 0.28 standard
deviations in numeracy at private other Christian schools, and 0.32 standard deviations
in reading and 0.52 standard deviations in numeracy at non-Christian private faith
schools, compared to 0.02 and 0.12 standard deviations among public school students.
Table 3.1: Selected school characteristics, by school type
(1) (2) (3) (4) (5) (6)
Public Private
All Catholic Christian Other faith Prep
No. of schools 554 95 33 39 7 15 Funding group 1 . 76 31 30 5 9 Funding group 2 . 19 2 9 2 6 Grade 4 only 101 1 0 0 0 1 Grade 7 only 35 4 0 2 1 1 Grades 4 and 7 418 85 33 37 6 13
Notes: see text and Data Appendix for details of sample selection and construction, and for variable definitions.
The remaining rows of Table 3.2 demonstrate the extent to which sorting across
schools produces differences in the observables characteristics of students attending
each school type. These patterns demonstrate the substantial potential for differences in
student characteristics to account for the observed differences in mean achievement
across school types. Students enrolled in private prep schools are the most positively
selected with respect to several observable characteristics that have a known
association with test scores. These students on average live in neighborhoods with very
high mean family income and highly educated household heads. Relatively few prep
schools students are Aboriginal or speak Punjabi at home, characteristics that in British
Columbia are associated with low mean test scores on average (see Friesen and Krauth
63
2011). Students enrolled in private non-Christian faith schools also live in relatively high
SES neighborhoods. However, these students are far less likely to speak English at
home than any other group. Notably, over one-third of these students are Punjabi-
speakers who attend Sikh faith schools. Students enrolled in Catholic schools are also
positively if slightly less strongly selected with respect to neighborhood SES. They are
more likely to speak English at home than public schools students and are less likely to
be Aboriginal. In contrast, neighborhood SES is very similar for students enrolled at
other Christian private schools compared to public school students. Punjabi speakers
and Aboriginal students are underrepresented at these schools, but the proportion that
speaks another non-English home language (e.g. Korean, Vietnamese or Tagalog) is
higher than among public school students.
Table 3.2: Selected student characteristics, grade 7 students with non-missing test scores, by school type
(1) (2) (3) (4) (5) (6)
Public Private
All Catholic Christian Other faith Prep
No. of students 87113 11924 4565 4448 532 2379 % of sample 88.00 12.00 4.60 4.50 0.50 2.40 Reading score 0.02 0.45 0.43 0.27 0.33 0.86 Numeracy score 0.12 0.52 0.54 0.28 0.52 0.93 Home language
English 0.67 0.74 0.81 0.72 0.36 0.71 Chinese 0.12 0.08 0.05 0.09 0.00 0.12 Punjabi 0.08 0.03 0.00 0.01 0.36 0.03
Other 0.13 0.16 0.14 0.18 0.29 0.15 Aboriginal 0.05 0.01 0.01 0.01 0.00 0.00 Female 0.48 0.49 0.52 0.46 0.54 0.48 Neighborhood mean
Immigrant 0.07 0.06 0.07 0.05 0.10 0.07 High school 0.25 0.25 0.25 0.26 0.24 0.23
Some college 0.30 0.28 0.29 0.30 0.24 0.24 Bachelor’s 0.17 0.22 0.20 0.15 0.23 0.36
Family income 6.71 7.89 7.06 6.67 7.07 12.00
Notes: see text and Data Appendix for details of sample selection and construction, and for variable definitions.
64
Table 3.3: Grade 7 student characteristics, by school type and mover status
ALL MOVERS (1) (2) (3) (4)
Public/public Public/private Private/public Private/private
No. of students 32370 1691 1410 862
% of full sample 32.68 1.71 1.42 0.87
Reading score -0.07 0.41 0.02 0.44
Numeracy score -0.01 0.53 0.07 0.45
Home language
English 0.69 0.59 0.61 0.69
Chinese 0.09 0.17 0.05 0.09
Punjabi 0.07 0.06 0.17 0.03
Other 0.15 0.20 0.16 0.20
Aboriginal 0.08 0.02 0.04 0.00
Female 0.48 0.44 0.47 0.47
Neighborhood mean
Immigrant 0.07 0.07 0.07 0.06
High school 0.26 0.24 0.25 0.24
Some college 0.31 0.28 0.29 0.28
Bachelor’s 0.15 0.23 0.18 0.24
Family income 6.39 8.21 6.98 8.43
COMPULSORY MOVERS ONLY
Public/public Public/private Private/public Private/private
No. of students 14946 179 109 63
% of full sample 15.09 0.18 0.11 0.06
Reading score -0.07 0.33 0.06 0.54
Numeracy score -0.04 0.42 -0.03 0.60
Home language
English 0.77 0.61 0.75 0.90
Chinese 0.08 0.15 0.07 0.02
Punjabi 0.04 0.07 0.04 0.02
Other 0.12 0.20 0.14 0.06
Aboriginal 0.07 0.03 0.04 0.02
Female 0.47 0.48 0.45 0.49
Neighborhood mean
Immigrant 0.06 0.08 0.05 0.04
High school 0.26 0.25 0.25 0.23
Some college 0.32 0.29 0.32 0.28
Bachelor’s 0.15 0.18 0.14 0.30
Family income 6.43 6.74 6.48 11.49
65
Notes: see text and Data Appendix for details of sample selection and construction, and for variable definitions.
The variation across school types in gender composition is also worth noting.
Private prep school students are 48 percent female, the same proportion as in public
schools. Catholic schools and other (non-Christian) faith private schools are
disproportionately female (52 percent and 54 percent respectively), while other (non-
Catholic) Christian private schools are disproportionately male (46 percent female).
Table 3.3 presents characteristics of students who change schools between
grades 4 and 7 by type of move. This mobility is used to identify school and student fixed
effects in the student fixed effects specification of our model. As described above,
identification of school effects requires that mobility patterns between private and public
schools are uncorrelated with unobserved factors that affect test score growth between
grades 4 and 7, conditional on time-varying student characteristics. The second and
third columns of Table 3.3 present characteristics of students who move from a public
school in grade 4 to a private school in grade 7 and vice versa. Among all movers,
shown in the top panel, the characteristics of students who move from public to private
schools differ on average from those who move from private to public schools in a
number of dimensions. To the extent that unobservable characteristics that affect test
score growth are correlated with these observables, this evidence suggests that the
student FE estimator may be subject to bias from this source.
The lower panel of Table 3.3 presents characteristics of students who change
schools, by type of move, within the sub-sample of students who were forced to change
schools between grades 4 and 7 because of the grade configuration of their grade 4
schools. Relatively few movers in this group move between the private and public
sectors. Of those who do, students who move from public to private schools reside in
neighborhoods with very similar mean family income to those who move from private
schools to public schools, and their test scores are more similar to one another. This
66
pattern of observables suggests that estimates from this sample may be less prone to
bias from this source.
3.5.2. Results from the private school indicator model
The main results from the simple private school indicator version of our model
are reported in Table 3.4. The top panel reports results for reading and the bottom panel
for numeracy. The first two columns report results from the value-added model. Along
with individual covariates, the specification reported in the first column includes a set of
neighborhood characteristics, while the second column includes postal code fixed
effects. The results show that the estimates of the private school effect is not sensitive to
these alternative methods of controlling for unobserved factors that are correlated with
neighborhood characteristics. These estimates are positive and statistically significant
for both reading and numeracy, with effect sizes between 0.16 and 0.18 standard
deviations.
The remaining columns of table 3.4 present results from the corresponding
student fixed effects model. The estimated effect for the full sample, presented in the
third column, is positive and statistically significant, but substantially smaller than in the
value-added specification, only 0.10 standard deviations in reading and 0.08 standard
deviations in math. The relatively small magnitude of these estimates compared to the
value-added estimates could reflect several factors. First, they incorporate the effects of
middle schools that don’t offer grade 7, while the value-added estimates do not. Second,
the effects of private schools on students who don’t change school sectors may be
greater than for those who do; since stayers do not contribute to the identification of
school fixed effects, this could result in lower estimates.
67
Table 3.4: Estimates of private school effect on reading and numeracy scores, value-added and student fixed effects models
Notes: Standard errors clustered at the school level. Additional control variables in all specifications include indicators for English as a Second Language and Aboriginal identity. Census neighborhood controls consist of mean family income in the student’s Census EA/DA, proportion of household heads in the student’s Census EA/DA who are immigrants, whose highest level of education is high school, a trade certificate, some college and a bachelor’s degree, and.
aDependent variable is the student’s grade 7 FSA test score. Additional control
variables include gender, home language (Chinese, Punjabi, other non-English) and year fixed
effects. b
Dependent variable is the student’s FSA test score in grade g. Additional control variables include grade-by-year fixed effects.
As described earlier, it is also possible that the student fixed effects estimates
are subject to bias from dynamic selection into mover status. We address this threat by
estimating our model for samples that exclude students who move between school
(1) (2) (3) (4) (5) (6)
READING
Value-addeda Student Fixed Effectsb
No movers between Compulsory
movers All students All students G4/G5 G6/G7
Private school 0.17*** 0.16*** 0.10*** 0.09*** 0.12*** 0.12**
[0.022] [0.018] [0.017] [0.020] [0.019] [0.036]
R-squared 0.484 0.502 0.028 0.027 0.027 0.047 # of students 84774 84774 169419 151950 149071 26450 Census EA characteristics x x x x x Postal code fixed effects x
NUMERACY
Value-addeda Student Fixed Effectsb
No movers between Compulsory
movers All students All students G4/G5 G4/G5
Private school 0.18*** 0.17*** 0.08*** 0.06** 0.10*** 0.09*
[0.035] [0.028] [0.021] [0.023] [0.023] [0.045]
R-squared 0.471 0.494 0.056 0.054 0.056 0.089 # of students 83375 83375 166613 149573 146799 25843 Census EA characteristics x x x x x Postal code fixed effects x
68
sectors after different grades. The results in the fourth column of Table 3.4 exclude
students who change school sectors immediately after grade 4; this exclusion will
eliminate bias from transitory shocks that systematically affect the school sector choice
of these movers and are correlated with grade 4 test scores. The results in the fifth
column exclude students who change school sectors immediately after grade 6; this
exclusion will eliminate bias from transitory shocks that systematically affect the school
sector choice of these movers and are correlated with grade 7 test scores. Excluding
post-grade 4 movers produces a smaller point estimate while excluding post-grade 6
movers produces a larger point estimate. This pattern of results is consistent with a
scenario where negative shocks to test scores are more highly correlated with transitory
shocks that cause public school students to move to private schools than with transitory
shocks that cause private school students to move to public schools. The results in final
column of table 4 correspond to the sub-sample of students who are required to change
schools between grades 4 and 7 because of the grade configuration of their grade 4
schools. The point estimates in both reading and numeracy are very close to those from
the full sample, despite the fact that they reflect effects for a sub-set of schools.
3.5.3. Results from the individual school fixed effects model
We next present estimates from specifications that include a full set of school
fixed effects rather than a simple private school indicator variable. These estimates allow
us to investigate heterogeneity within the public and private school sectors.15 We begin
this investigation by computing the differences between the average of the estimated
school fixed effects for each type of private school and the average for public schools.
These enrolment-weighted mean differences are presented in table 3. 5. As in table 3.4,
the
__________________________________ 15 Recent work on charter school quality uses population data on public schools to investigate
within-sector heterogeneity and characterize the evolution of the distribution of charter school effects over time in Texas (Baude et al. 2015) and North Carolina (Ladd et al. 2015). This dynamic perspective is of less interest in the context of B.C.’s mature private school sector.
69
Table 3.5: Student-weighted means of estimated private school fixed effects on reading and numeracy scores, by type of private school, value-added and student fixed effects models
Notes: Standard errors clustered at the school level. Reported effects are differences between student-weighted averages of estimated private school and public school fixed effects, by private school type. Control variables in all specifications include indicators for English as a Second Language and Aboriginal identity. Census neighborhood controls consist of mean family income in the student’s Census EA/DA, proportion of
(1) (2) (3) (4) (5) (6)
READING
Student Fixed Effectsb
Value-addeda No movers between Compulsory
movers All students All students G4/G5 G6/G7
Prep 0.26*** 0.25*** 0.17*** 0.20*** 0.14** 0.34***
[0.064] [0.046] [0.036] [0.034] [0.044] [0.083]
Catholic 0.21*** 0.21*** 0.11*** 0.09*** 0.15*** 0.04
[0.030] [0.022] [0.022] [0.026] [0.026] [0.079]
Other Christian 0.08*** 0.07*** 0.03 -0.01 0.05* 0.07
[0.021] [0.020] [0.022] [0.025] [0.027] [0.049]
Other faith 0.26** 0.25** 0.20*** 0.20** 0.23*** 0.59***
[0.086] [0.080] [0.060] [0.064] [0.067] [0.149] R-squared 0.485 0.504 0.028 0.027 0.027 0.048 # of students 84774 84774 169419 151950 149071 26450 Census char. x x x x x Postal code FE x
NUMERACY
Student Fixed Effectsb
Value-addeda No movers between Compulsory
movers All students All students G4/G5 G6/G7
Prep 0.26** 0.26** 0.18*** 0.18*** 0.16** 0.35***
[0.087] [0.061] [0.047] [0.043] [0.055] [0.090]
Catholic 0.29*** 0.29*** 0.14*** 0.13*** 0.16*** 0.18*
[0.046] [0.035] [0.027] [0.031] [0.031] [0.075]
Other Christian 0.04 0.00 -0.05 -0.11*** -0.02 0.02
[0.032] [0.032] [0.025] [0.026] [0.027] [0.059]
Other faith 0.24* 0.25* 0.18* 0.20** 0.24** -0.17
[0.119] [0.089] [0.071] [0.064] [0.074] [0.178] R-squared 0.473 0.497 0.056 0.055 0.056 0.090 # of students 83375 83375 166613 149573 146799 25843 Census EA char. x x x x x Postal code FE x
70
household heads in the student’s Census EA/DA who are immigrants, whose highest level of education is high school, a trade certificate, some college and a bachelor’s degree. aDependent variable is the student’s grade 7 FSA test score. Additional control variables include grade 4 test score, gender, home language (Chinese, Punjabi, other non-English) and year fixed effects. bDependent variable is the student’s FSA test score in grade g. Additional control variables include grade-by-year fixed effects and a full set of individual fixed effects.
first two columns present results from the value-added model, with neighborhood
characteristics and postal code fixed effects respectively, and the remaining columns
present results from the student fixed effects model estimated for several different
samples.
The mean differences for the value-added school effects estimates are positive
and statistically significant for all private school types. The effect sizes are similar for
prep schools, Catholic schools and non-Christian faith schools, ranging between 0.21
and 0.29 standard deviations. The effect size for Other (non-Catholic) Christian private
schools is smaller, 0.08 standard deviations in reading and 0.04 standard deviations and
statistically insignificant in numeracy. The corresponding mean differences for the
student fixed effects model are smaller for non-Christian faith schools and prep schools,
but are still positive, substantial and statistically significant, ranging between 0.17 and
0.20 standard deviations. The point estimates for Catholic schools are smaller, 0.11
standard deviations in reading and 0.14 standard deviations in numeracy, but remain
statistically significant. The point estimates for other Christian schools are small and
statistically insignificant. In most cases, excluding post-grade 4 movers produces a
smaller point estimate while excluding post-grade 6 movers produces a larger point
estimate. Again, this pattern of results is consistent with a scenario where negative
shocks to test scores are more highly correlated with factors that cause public school
students to move to private schools than with factors that cause private school students
to move to public schools. In the case of prep schools, this pattern is reversed. When we
restrict the sample to compulsory movers only, the effects for other Christian schools
remain small and statistically significant. Among Catholic schools, the effect for reading
is now smaller and statistically insignificant; among other faith schools the effect for
71
numeracy is now negative and statistically insignificant. The estimated effects for the
subset of prep schools included in this sample remain positive, substantial and
statistically significant.
We further characterize heterogeneity within and across school types in table
3.6, using estimates of individual school and student fixed effects from the student fixed
effects model. The first column shows the differences between school-weighted
averages of estimated private school and public school fixed effects, by private school
type. These mean estimates differ somewhat from those presented in the third column of
table 3.5 because they are school-weighted, rather than student-weighted, averages of
estimated school fixed effects. The second column of table 3.6 shows differences
between average estimated student fixed effects in private and public schools, by private
school type. Prep schools enroll the highest achieving students; these students on
average would score about half a standard deviation above the public school mean in
both reading and numeracy if they were enrolled in public schools. Students enrolled in
Catholic and other Christian private schools are also positively selected on ability, but to
a lesser degree. These students would score between 0.14 and 0.22 standard
deviations above the public school mean if they were enrolled in public schools.
Students attending other faith private schools are only slightly positively selected on
ability, by 0.07 to 0.06 standard deviations.
The last two columns of table 3.6 characterize the standard deviations of
estimated school and student effects by type of school. The standard deviation of school
effects is approximately half again as large among private schools compared to public
schools. The prep, Catholic and especially other faith private school sectors are all
substantially more heterogeneous with respect to school effects than public schools,
while other Christian private schools are no more heterogeneous than public schools.
The within-school variation in student effects is moderately lower on average among
prep schools and other faith private schools compared to public schools. These schools
may be attracting a more homogeneous group of applicants, or applying more selective
admission procedures. Students at Catholic and other Christian private schools exhibit
slightly higher degrees of heterogeneity on average, but these schools still enroll
72
relatively homogeneous students with respect to reading ability compared to public
schools.
Figure 1 presents kernel density estimates of the distribution of school fixed
effects for private and public schools estimated from the student FE model. The range of
estimated school fixed effects is wider for private than for public schools because of
fatter tails at both ends of the distribution. A fairly small proportion of private schools are
less effective than the average public school in reading. This proportion is higher in the
case of numeracy. Figure 2 presents analogous kernel density estimates of the student-
weighted distribution of school fixed effects, which leads to a similar conclusion. Figure 3
presents estimates of the distribution of student fixed effects from the same set of
estimates. The entire distribution of student fixed effects is shifted slightly to the right in
private schools compared to public schools.
73
Table 3.6: School-weighted means and standard deviations of estimated private school and student fixed effects on reading and numeracy scores, by type of school, student fixed effects model
Notes: Standard errors are bootstrapped. Dependent variable is the student’s FSA test score in grade g.
Control variables include indicators for English as a Second Language, mean family income in the
student’s Census EA/DA, proportion of household heads in the student’s Census EA/DA who are
immigrants, whose highest level of education is high school, a trade certificate, some college and a
bachelor’s degree, grade-by-year fixed effects. Reported effects are: adifferences between school-weighted
averages of estimated private school and public school fixed effects, by private school type; bstandard
deviations of estimated individual school fixed effects, by school type; and cwithin-school standard
deviations of estimated individual student fixed effects, by school type.
(1) (2) (3) (4)
READING
Mean difference from public
school meana
Standard deviation
School
effect/se
Student
effect/se
School
effectb
Student effect
(within school)c
All public 0.17 0.79
All private 0.10*** 0.27*** 0.30 0.75
0.021 0.023
Prep 0.22*** 0.57*** 0.23 0.71
0.037 0.044
Catholic 0.11*** 0.22*** 0.21 0.76
0.033 0.035
Other Christian 0.03 0.21*** 0.23 0.77
0.034 0.033
Other faith 0.14** 0.07 0.76 0.74
0.054 0.060
NUMERACY
Mean difference from public
school meana Standard deviation
School
effect/se
Student
effect/se
School
effectb
Student effect
(within school)c
All public 0.23 0.81
All private 0.09*** 0.22*** 0.31 0.80
0.019 0.020
Prep 0.21*** 0.51*** 0.36 0.77
0.036 0.040
Catholic 0.15*** 0.14*** 0.26 0.80
0.035 0.039
Other Christian -0.06* 0.18*** 0.27 0.81
0.030 0.034
Other faith 0.21*** 0.06 0.45 0.74
0.066 0.065
74
Figure 3.1: Kernel density estimates of the distribution of estimated school effects, student FE model, full sample, by school sector Reading
Numeracy
75
Figure 3.2: Kernel density estimates of the student-weighted distribution of estimated school effects, student FE model, full sample, by school sector Reading
Numeracy
76
Figure 3.3: Kernel density estimates of the distribution of estimated student effects, student FE model, full sample, by school sector Reading
Numeracy
77
3.6. CONCLUSION
All of the results in this paper point to the relative success of the average private
primary school in British Columbia with respect to literacy and numeracy skill
development, compared to its public counterpart. This evidence contradicts the results of
most previous studies of the effects of private schools on test scores in other advanced
economies; Catholic private schools in particular have been associated with no better or
even relatively poor test scores in both the U.S. and Australia (e.g. Altonji et al. 2005;
Elder and Jepsen 2014; Nghiem et al. 2015). This qualitative difference in our
conclusions about the relative effectiveness of private schools may reflect differences in
institutions. Like private schools in many jurisdictions, B.C.’s private schools enjoy
substantial autonomy with respect to teacher hiring, firing, discipline and remuneration,
are free to apply a wide range of admissions criteria and can charge tuition. Unlike
private schools in some jurisdictions, B.C. private schools are required to hire teachers
who are provincially certified, and they receive partial funding from the provincial
government. Published school-level test score results include private schools, and these
receive a substantial amount of media and public attention.
Nevertheless, the range of quality among private schools is substantial, and
weaker private schools appear to be less effective than many public schools. The
survival of these schools, many of which have been in operation for decades, invites
several potential explanations. One possibility is that parents may value private schools
for a variety of reasons beyond their contributions to cognitive skill development;
religious or moral instruction, peer quality, enriched supervision and monitoring of
student behavior are examples of school qualities that parents are willing to pay for that
may not substantially affect cognitive skills. A second possibility is that some private
schools may be located in neighborhoods with relatively weak public schools. If it is
cheaper for some parents to send their children to a local private school than to relocate
to a more expensive neighborhood with better public schools, these private schools
78
need only outperform their local public counterparts. To the extent that the equilibrium
distribution of school quality reflects these sorts of constraints on individual school
choice, it also suggests a possible explanation for why our results differ qualitatively for
the average private school than those of previous studies. Unlike research based on
national survey data, we are able to include detailed geographic controls in our
population-based study. Our estimates therefore are based to a greater degree on
comparisons of public and private schools that enroll students who live in similar
neighborhoods.
While we cannot claim to have fully addressed the issue of non-random selection
into private schools, we also have several reasons to believe that any positive bias from
this source may be offset at least to some degree by other sources of negative bias. In
the value-added estimates, measurement error in the lagged test score may be
substantial, leading to attenuation bias in our estimates of private school effects.
Moreover, our approach is more likely to produce estimates that capture general
equilibrium effects of private school competition on public school quality compared to
previous studies, since these effects will be more pronounced in comparisons among
schools that are geographically proximate. Since, these spillover effects onto public
school quality via competition will reduce the quality difference between public and
private schools, our estimates will capture only part of the total effect of private schools
on student achievement (see Epple et al. 2015 for a review of relevant evidence of the
effects of private school competition).
Our methodology provides no direct evidence of the mechanisms that may be
driving our results. However, it is of some interest to note that our analysis by private
school type finds positive effects for some but not all groups of faith private schools as
well as positive effects for secular prep private schools. This pattern suggests that it is
not faith-based education per se that is contributing to the relative effectiveness of some
private schools. This interpretation is consistent with existing direct evidence that faith
public schools are no more effective than secular public schools (Gibbons and Silva
2011). It is also worth noting that, while private schools on average enroll higher quality
students and deliver higher quality outcomes, this is not true in all private school sectors.
79
In particular, other Christian schools enroll relatively high quality students but do not get
better results on average than public schools, while other faith schools enroll fairly
average students but get better than average results. This pattern is not consistent with
a simple story of peer effects driving variation in school performance across public
versus private school types.
80
References
Abbasi, H. Farjadi, Gh. (1999) Labour Force Participation in Iran. (In Farsi) In: ed. M. Tabibian. Iran’s Economy 2. Institute for Research on Planning and Development Publication, Tehran, Iran.
Abbasi-Shavazi, M. J., McDonald, P., & Hosseini-Chavoshi, M. (2009). The fertility transition in Iran: Revolution and reproduction. Springer, Dordrecht
Abowd, J., R. Creecy and F. Kramarz (2002). Computing person and firm effects using linked longitudinal employer-employee dataset, Technical report, US Census Bureau.
Aghajanian, A. (1995). A new direction in population policy and family planning in the Islamic Republic of Iran. Asia-Pacific Population Journal / United Nations, 10(1), 3–20.
Agüero, J. M., & Marks, M. S. (2008). Motherhood and female labour force participation: evidence from infertility shocks. The American Economic Review, 500–504.
Alderman, Harold, Peter F. Orazem and Elizabeth M. Paterno (2001). School quality, school cost and the public/private school choices of low-income households in Pakistan. Journal of Human Resources 36(2): 304-26.
Almond, D., & Edlund, L. (2008). Son-biased sex ratios in the 2000 United States Census. Proceedings of the National Academy of Sciences, 105(15), 5681–5682.
Altonji, Joseph G., Todd E. Elder, and Christopher R. Taber (2005a). An evaluation of instrumental variable strategies for estimating the effects of Catholic schooling,Journal of Human Resources 40(4): 791-821.
Altonji, Joseph G., Todd E. Elder, and Christopher R. Taber (2005b). Selection on observed and unobserved variables: Assessing the effectiveness of Catholic schools, Journal of Political Economy 113: 151–184.
Andrabi, T., J. Das, A.I. Khwwja and T. Zajonc (2011). Do value-added estimates add value? Accounting for learning dynamics. American Economic Journal: Applied Economics 3(3): 29-54.
Angrist J, Bettinger E, Kremer M. (2006). Long-term educational consequences of secondary school vouchers: evidence from administrative records in Colombia, American Economic Review 96(3):847-62.
Angrist Joshua, Eric Bettinger E, Bloom E, King E, Kremer M. (2002). Vouchers for private schooling in Columbia: randomized natural experiment. American Economic Review 92(5):1535-58.
81
Angrist, J. D., & Evans, W. N. (1998). Children and Their Parents’ Labour Supply: Evidence from Exogenous Variation in Family Size. The American Economic Review, 88(3), 450–477.
Angrist, Joshua, P.A. Pathak and C.R. Walters (2013). Explaining charter school effectiveness. American Economic Journal: Applied Economics 5(4): 1-27.
Azimi, E. "The effect of children on female labour force participation in urban Iran." IZA Journal of Labour & Development 4.1 (2015): 1-11.
Baude, P.L., M. Casey, E.A. Hanushek and S.G. Rivkin (2014). The evolution of charter school quality. NBER Working Paper Number 20645, National Bureau of Economic Research.
Becker GS, Lewis HG (1974) Interaction between quantity and quality of children. In Economics of the family: Marriage, children, and human capital. UMI, pp 81–90
Ben-Porath, Y., & Welch, F. (1976). Do sex preferences really matter? The Quarterly Journal of Economics, 285–307.
Bhalotra, S., & Attfield, C. (1998). Intrahousehold resource allocation in rural Pakistan: a semiparametric analysis. Journal of Applied Econometrics, 13(5), 463–480.
Blundell, R., Kristensen, D., & Matzkin, R. L. (2013). Control functions and simultaneous equations methods. The American Economic Review, 103(3), 563-569.
British Columbia Federation of Independent Schools Association (2015). Who are we? http://www.fisabc.ca/who-are-we/history
British Columbia Ministry of Education (2001) Interpreting and communicating Foundation Skills Assessment Results 2001 (Victoria: Government of British Columbia).
British Columbia Ministry of Education (2005). Overview of Independent Schools in British Columbia. http://www.bced.gov.bc.ca/independentschools/geninfo_05.pdf
British Columbia Ministry of Education (2011). Provincial Report: Student Statistics. http://www.bced.gov.bc.ca/reports/pdfs/student_stats/prov.pdf
Brochu, Pierre, Marie-Anne Deussing, Koffi Houme and Maria Chuy (2012). Measuring up: Canadian results of the OECD PISA study. Council of Ministers of Education, Canada.
Bronars, S. G., & Grogger, J. (1994). The Economic Consequences of Unwed Motherhood: Using Twin Births as a Natural Experiment. American Economic Review, 84(5), 1141–1156.
Browning, M., Chiappori, P.-A., & Lewbel, A. (2013). Estimating Consumption Economies of Scale, Adult Equivalence Scales, and Household Bargaining Power. Review of Economic Studies, 80(4), 1267–1303.
Buddin, Richard J., Joseph J. Cordes and Sheila Nataraj Kirby (1998). School choice in California: Who chooses private schools? Journal of Urban Economics 44(1): 110-134
82
Carbonaro, William (2006). Public-private differences in achievement among kindergarten students: differences in learning opportunities and student outcomes. American Journal of Education 113(1): 31-66.
Chiappori, P.-A. (1988). Rational Household Labour Supply. Econometrica, 56(1), 63.
Chiappori, P.-A. (1992). Collective Labour Supply and Welfare. Journal of Political Economy, 100(3), 437–67.
Chun H, Oh J (2002) An instrumental variable estimate of the effect of fertility on the labour force participation of married women. Applied Economics Letters 9(10):631-634
Clark, D. (2009). The performance and competitive effects of school autonomy.Journal of Political Economy, 117(4):745-783.
Cowley, Peter and Stephen T. Easton (2003). Report Card on British Columbia’s Elementary Schools: 2003 Edition (Vancouver, B.C.: Fraser Institute).
Das Gupta, M. Zhenghua, J. Bohua, L. Zhenming, X. Chung, W. & Hwa-Ok, B. Why is Son Preference so Persistent in East and South Asia? A Cross-Country Study of China, India, and the Republic of Korea. (2003). The Journal of Development Studies, 40(2), 153-187.
Deaton, A. (1997). The Analysis of Household Surveys: A Microeconometric Approach to Development Policy. World Bank Publications.
Deming, Dave J., Justine S. Hastings, Thomas J. Kane and Douglas O. Staiger (2014). School choice, school quality and academic achievement. American Economic Review 104(3): 991-1013.
Dunbar, G. R., Lewbel, A., & Pendakur, K. (2013). Children’s Resources in Collective Households: Identification, Estimation, and an Application to Child Poverty in Malawi. American Economic Review, 103(1), 438–471.
Elder, Todd and Christopher Jepsen (2014). Are Catholic primary schools more effective than public primary schools? Journal of Urban Economics 80:28-38.
Epple, Dennis, Richard E. Romano and Miguel Urquiola (2015a). School vouchers: A survey of the economics literature. NBER Working Paper Number 21523, National Bureau of Economic Research.
Evans, William N. and Robert M. Schwab (1995). Finishing high school and starting college: Do Catholic schools make a difference? Quarterly Journal of Economics 110 (4): 941–974.
Figlio, David N. and Joe A. Stone (1997). School choice and student performance: Are private schools really better? Discussion Paper 1141-97, Institute for Research on Poverty.
Figlio, David N., and Cassandra M. D. Hart (2014). Competitive effects of means-tested school vouchers. American Economic Journal: Applied Economics 6(1): 133-56.
Fraser Institute (2008). Who We Are. http://www.fraserinstitute.org/aboutus/whoweare/
83
Friesen, Jane and Brian V. Krauth (2011). Ethnic enclaves in the classroom. Labour Economics 18: 656-663.
Gibbons, Stephen and Olmo Silva (2011). Faith primary schools: Better schools or better pupils? Journal of Labour Economics 29(3): 589-635.
Goldin, C. (1995). Career and family: College women look to the past. (No. w5188)
National Bureau of Economic Research.
Gong, X., van Soest, A., & Zhang, P. (2000). Sexual Bias and Household Consumption: A Semiparametric Analysis of Engel Curves in Rural China (IZA Discussion Paper No. 212). Institute for the Study of Labour (IZA).
Grogger, Jeffrey and Derek Neal (2000). Further evidence on the effects of Catholic secondary schooling. In: Gale,William G.,Pack, Janet Rothenberg (Eds.), Brookings-Wharton Papers On Urban Affairs, 2000. Brookings Institution Press,Washington, DC: 151–193.
Gronau, R. (1988). Consumption Technology and the Intrafamily Distribution of Resources. Journal of Political Economy, 96(6), 1183–1205.
Gugl, E., & Welling, L. (2012). Time with sons and daughters. Review of Economics of the Household, 10(2), 277-298.
Hanushek, Eric A., Kain, John F., Rivkin, Steve G.(2004). Disruption versus Tiebout improvement: the costs and benefits of switching schools. Journal of Public Economics 88(9–10): 1721–1746.
Heckman, J. J. and R. Robb (1985), “Alternative Methods for Evaluating the Impact of Interventions,” in Longitudinal Analysis of Labour Market Data, (eds J. Heckman and B. Singer), CUP.
Hill, M. A., & King, E. (1995). Women’s education and economic well-being. Feminist Economics, 1(2), 21–46.
Hoodfar, H. (1996). Bargaining with fundamentalism: women and the politics of population control in Iran. Reproductive Health Matters, 4(8), 30–40.
Hurst, E. and A. Lusardi (2004): “Liquidity Constraints, Household Wealth, and Entrepreneurship,” Journal of Political Economy, 112, 319–347.
Imberman, Scott A. (2011). Achievement and behavior in charter schools: Drawing a more complete picture. Review of Economics and Statistic 93(2): 416-435.
Jacob, Brian A., Lars Lefgren and David P. Sims (2010). The persistence of teacher-induced learning. Journal of Human Resources 45(5): 915-943.
Jepsen, Christopher (2003). The effectiveness of Catholic primary schooling. Journal of Human Resources 38(4): 928-941.
Kane, Thomas J. and Douglas O. Staiger (2008). Estimating teacher impacts on student achievement: an experimental evaluation. Working Paper 14607, National Bureau of Economic Research.
84
Kane, Thomas J., Daniel F. McCaffrey, Trey Miller and Douglas O. Staiger (2013). Have we identified effective teachers? Validating measures of effective teaching using random assignment. Research Paper. MET Project. Technical report. Bill and Melinda Gates Foundation.
Kingdon, G. G. (2005). Where has all the bias gone? Detecting gender bias in the intrahousehold allocation of educational expenditure. Economic Development and Cultural Change, 53(2), 409–451.
Kramarz, Francis, Stephen Machin and Amine Ouazad (2014). Using compulsory mobility to identify school quality and peer effects. Oxford Bulletin of Economics and Statistics :1-22.
Ladd, Helen F., Charles T. Clotfelter and John B. Holbein (2015). The growing segmentation of the charter school sector in North Carolina. NBER Working Paper No. 21078. National Bureau of Economic Research.
Lancaster, G., Maitra, P., & Ray, R. (2003). Endogenous power, household expenditure patterns and new tests of gender bias: evidence from India. In Preliminary Version of paper presented at the NEUDC conference.
Lefebvre, Pierre, Philip Merrigan and Matthieu Verstraete (2011). Public subsidies to private schools do make a difference for achievement in mathematics: Longitudinal evidence from Canada. Economics of Education Review 30: 79-98.
Lise, J., & Seitz, S. (2011). Consumption inequality and intra-household allocations. The Review of Economic Studies, 78(1), 328-355.
Long, X. (1989). The higher sex ratio in the modern population of China. Wei Ding Gao, 8, 53-59.
Lubienski, Christopher, Corinna Crane and Sarah Thule Lubienski (2008) What do we know about school effectiveness? Academic gains in public and private schools. Phi Delta Kappan 89(9): 689-695.
Majbouri, M. (2010). Against the Wind: Labour Force Participation of Women and Economic Instability in Iran. Available at SSRN 2419323.
Majbouri, M. (2015). Oil, Patriarchy, and Female Labour Force Participation. Working paper
Martinez-Mora, Francisco (2006). The existence of non-elite private schools. Journal of Public Economics 90: 1505-1518.
McEwan, Patrick J. (2001). The effectiveness of public, Catholic and non-religious private schools in Chile’s voucher system. Education Economics 9(2): 103-128.
Mehryar, A. H., Ahmad-Nia, S., & Kazemipour, S. (2007). Reproductive health in Iran: pragmatic achievements, unmet needs, and ethical challenges in a theocratic system. Studies in Family Planning, 38(4), 352–361.
85
Mehryar, A. H., Farjadi, G., Tabibian, M., & Population Studies and Research Center for Asia and the Pacific (Iran). (2002). Labour force participation of women in Iran: trends and covariates. Tehran, Iran: Ministry of Science, Research & Technology, Population Studies and Research Center for Asia and the Pacific.
Michael R. T., 1974. "Education and the Derived Demand for Children," NBER Chapters, in: Economics of the Family: Marriage, Children, and Human Capital, pages 120-159 National Bureau of Economic Research, Inc.
Moghadam, V. M. (1992). Patriarchy and the Politics of Gender in Modernising Societies: Iran, Pakistan and Afghanistan. International Sociology, 7(1), 35–53.
Muellbauer, J. (1976). Community Preferences and the Representative Consumer. Econometrica, 44(5), 979–99.
Muralidharan, Karthik and Venkatesh Sundararaman (2015). The aggregate effect of school choice: Evidence from a two-stage experiment in India. Quarterly Journal of Economics, forthcoming.
Neal, Derek (1997). The effect of Catholic secondary schooling on educational attainment, Journal of Labour Economics 15(1 Part 1): 98–123.
Nghiem, Hong Son, Ha Trong Nguyen, Rasheda Khanam and Luke B. Connelly (2015). Does school type affect cognitive and non-cognitive development in children? Evidence from Australian primary schools. Labour Economics 33: 55-65.
Oster, Emily (2015). Unobservable selection and coefficient stability: Theory and evidence. Working Paper.
Polachek, D. E. and Polachek, S. W. (1989), An Indirect Test of Children's Influence on Efficiencies in Parental Consumer Behavior. Journal of Consumer Affairs, 23: 91–110.
Rasul-Ronning Z. (2013). Women in Iraq and Iran: Leadership and the Right to Equal Participation. In: Winkler, P. Confronting the International Patriarchy: Iran, Iraq and the United States of America. Newcastle upon Tyne, UK: Cambridge Scholars Publishing.
Reardon, Sean F., Jacob F. Cheadle and Joseph P. Robinson (2009). The effects of Catholic schooling on math and reading development in kindergarten through fifth grade. Journal of Research on Educational Effectiveness 2(1): 45-87
Rivkin, Steve G., Eric A. Hanushek and John F. Kain (2005). Teachers, schools, and academic achievement. Econometrica 73(2): 417–458.
Rose, E. (1999). Consumption Smoothing and Excess Female Mortality in Rural India. Review of Economics and Statistics, 81(1), 41–49.
Rosenzweig, M. R., & Wolpin, K. I. (1980). Testing the quantity-quality fertility model: the use of twins as a natural experiment. Econometrica: Journal of the Econometric Society, 48(1), 227–240.
Rothbarth, E. (1943), "Note on a Method of Determining Equivalent Income for Families of Different Composition," in: Madge, C. (Ed.), Wartime Pattern of Saving and Spending, London: Faber and Faber Limited, 123-130.
86
Rouse, Cecilia and Lisa Barrow (2009). School vouchers and student achievement: Recent evidence, remaining questions, Annual Review of Economics 17-42.
Sander, William (1996). Catholic grade schools and academic achievement. Journal of Human Resources 31(2): 540-548.
Schultz, TP (1981) Economics of population. Perspectives on Economics Series
Shack, Daniel (2015). The impact of parents’ uncertainty and learning on dynamic educational choices and student achievement. Presented at the CEA meetings.
Sharabi, Hisham. 1988. Neopatriarchy: A Theory of Distorted Change in Arab Society. New York: Oxford University Press.
Singh, Abhijeet (2015). Private school effects in urban and rural India: Panel estimates at primary and secondary school ages. Journal of Development Economics 113: 16-32.
Stock JH, Yogo M (2005) Testing for weak instruments in linear IV regression. Identification and inference for econometric models: Essays in honor of Thomas Rothenberg
Subramanian, S., & Deaton, A. (1991). Gender effects in Indian consumption patterns. Sarvekshana, 14(4), 1–12.
Todd, Petra E. and Kenneth I. Wolpin (2003). On the specification and estimation of the production function for cognitive achievement. Economic Journal 113(485): F3–F33.
UNICEF (2007) State of the World’s Children : the Double Dividend of Gender Equality, New York: UNICEF
United Nations Population Fund (UNFPA) (2010) State of the World’s Population 2010. From conflict and crisis to renewal: generations of change, UNFPA, New York, p.96
Working, H. (1943). Statistical Laws of Family Expenditure. Journal of the American Statistical Association, 38(221), 43–56.
Zimmermann, L. (2012). Reconsidering Gender Bias in Intrahousehold Allocation in India. Journal of Development Studies, 48(1), 151–163.
87
Appendix A. Further Statistics
Table A1: Fertility and FLFP in Iran during 1990–2004
1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004
21–35 year old women
number of children
3.26 3.29 2.99 2.91 2.89 2.86 2.59 2.46 2.43 2.38 2.16 2.05 2.01 1.96 1.89
more than 2 children
0.64 0.64 0.56 0.54 0.54 0.53 0.46 0.42 0.41 0.40 0.33 0.30 0.30 0.27 0.25
FLFP 0.12 0.12 0.15 0.14 0.12 0.15 0.13 0.13 0.12 0.14 0.14 0.12 0.13 0.14 0.13
observation 3802 3656 2564 2730 4816 7742 4183 4182 2983 4385 4664 4454 5584 4060 4297
36–50 year old women
number of children
3.98 4.02 3.98 3.93 3.85 3.97 3.75 3.70 3.62 3.69 3.56 3.52 3.41 3.28 3.19
more than 2 children
0.75 0.76 0.75 0.75 0.74 0.77 0.74 0.75 0.73 0.73 0.72 0.71 0.69 0.67 0.64
FLFP 0.13 0.13 0.17 0.15 0.14 0.16 0.13 0.14 0.14 0.15 0.14 0.13 0.13 0.14 0.14
observation 2689 2835 1829 2086 3995 6926 3850 3832 3065 4806 4344 4474 5366 3946 4184
21–35 year old women with two or more children where the oldest child is younger than 18 years
number of children
3.67 3.71 3.54 3.45 3.44 3.38 3.21 3.10 3.07 3.01 2.91 2.84 2.83 2.73 2.70
more than 2 children
0.74 0.74 0.70 0.68 0.68 0.66 0.62 0.58 0.57 0.56 0.51 0.47 0.49 0.45 0.43
FLFP 0.11 0.11 0.14 0.13 0.11 0.13 0.11 0.11 0.10 0.12 0.12 0.10 0.10 0.12 0.11
observation 3280 3150 2059 2182 3836 6201 3141 3032 2162 3161 3050 2789 3408 2477 2517
Source: Iranian Household Expenditure and Income Surveys
Table A2: Under-18 Population by Gender and the Sex Ratio in Iran
1996 Census 2006 Census
male 14,465,424 11,712,684
female 13,889,253 11,142,509
sex ratio 1.041 1.051
Source: Statistical Yearbook, Statistical Center of Iran
88
Appendix B. Merging Student information to Census characteristics
We incorporate students’ residential information to our data set by merging it to Canadian Census. This enables us to create a rich data set by including variables such as mean family income and the percentage of household heads in the student’s Census Enumeration/Dissemination area who are immigrants, who are without high school diploma, who have a high school diploma, a post-secondary certificate or a bachelor’s degree.
We merge students in 2006 school year to 2006 census, students in 2001-05 school years to 2001 Census, and students in 1999-2000 school years into 1996 Census. We use Postal Code Conversion Files (PCCF) to associate each postal code in student data set to Enumeration Area (EA)/ Dissemination Area (DA) codes in Census.
Education variables are reported differently in different Census years. The above categorization provides us with variables that are consistent and comparable for different years of census.
Please refer to Statistics Canada website for further information regarding the Canadian Census, area codes, and PCCF.